KnowMore Ventures Journal

The operator thesis: why the best AI startups are built by people who have run the industry

hello@knowmoreventures.com (Sai) — Wed, 01 Apr 2026 00:00:00 GMT

Closing essay: a year of backing ten AI companies across PropTech, EdTech, Legal AI, Enterprise AI, and Consumer Social AI has produced one consistent conclusion. Domain depth is not a nice-to-have. It is the moat.

We have spent the last eighteen months backing ten AI companies across five verticals: PropTech, EdTech, Enterprise AI, Legal AI, and — most recently — Consumer Social AI. Every company in the portfolio was founded or co-built by someone who had spent years inside the industry they are now automating. Kanzi.ai (No. 01) was built by people who ran Oracle EBS implementations for Fortune 500 clients. RealtyBlocks (No. 02) and realestateindia.ai (No. 05) were built by people who understood why the Indian property transaction was broken from the inside. FlowAI (No. 03) was built by engineers who had spent careers designing the enterprise back-office workflows they are now automating with agents. SparkLearn (No. 04) was built by educators who knew why standard curriculum software failed to hold a child's attention. memtra.ai (No. 06) was built by engineers who had watched language models fail in production because they could not remember yesterday's context. DeReal (No. 07) was built by people who understood that property discovery had become an attention problem as much as a data problem, and that the platform that could hold a buyer's focus long enough to create genuine conviction would win the transaction. ClearPass (No. 08) was built by people who had processed property KYC manually and knew exactly which document caused every transaction delay. walkthesite.com (No. 09) was built by people who had accompanied buyers on site visits and understood what questions a video could never answer. And Saanjh (No. 11) was built by people who had watched an entire generation of urban Indians cycle through swipe-based dating apps and emerge more exhausted than connected — and who understood that the AI layer, finally capable of genuine values-aligned curation, was the tool that could make restraint a viable consumer product. The pattern is not a coincidence. It is the thesis.

Why domain depth is the moat, not the exception

The consensus view in venture capital is that AI is a horizontal technology and that the best AI companies are therefore built by teams with AI capability, not industry experience. We think this is wrong for most application-layer AI, and our portfolio data reinforces the view. The companies that have moved fastest from prototype to paying customer are the ones where the founders did not need to spend six months learning the problem. They already knew which workflow was broken, which data source was authoritative, which buyer had budget authority, and which compliance requirement was non-negotiable. That prior knowledge compresses the prototype-to-product arc from twelve months to three.

Figure: A timeline of ten KnowMore portfolio companies from first prototype to first paying customer: median sixty-eight days. The shortest path in each case ran through a founder who already knew the domain.

The AI capability question is also less differentiated than it appears. Every serious application-layer AI company in 2026 is building on Claude, GPT-4o, or Gemini. The model is a commodity. What is not a commodity is the training data, the fine-tuned domain vocabulary, the curated evaluation harness, and the integration with the specific enterprise system the buyer uses. All of those require domain knowledge, not AI research capability. A team that can build a RAG system on Claude with a curated corpus of Oracle MetaLink documents and a decade of support ticket history will outperform a team with more AI capability but less Oracle knowledge in every head-to-head evaluation.

The AI companies that will define the next decade are not being built in research labs. They are being built by the operators who know where the bodies are buried — and who finally have the tools to fix them.
— Sai, Managing General Partner, KnowMore Ventures, closing essay, April 2026

What the portfolio tells us about the next decade

The India–USA corridor is an underappreciated source of operator-founders: engineers and executives who have lived in both markets and understand where the friction points are on both sides.
The Studio co-build model — equity-for-engineering — is the fastest way to get an operator-founder to their first paid customer. Most operator-founders are strong on domain and distribution; they need technical velocity, not advice.
The best AI companies in regulated verticals (legal, compliance, healthcare) will be the ones that built their compliance posture into the prototype, not retrofitted it at the Series A.
The evaluation harness — a curated library of domain-specific test cases — is the compound moat. Build it early, grow it with every customer, share it with no one.

The closing observation from this first year of backing is about speed. Every company in the portfolio took longer to find its first design partner than we expected, and shorter to close its first paying customer once it had one. The design partner phase is a trust problem: enterprise buyers are cautious about AI tools from companies they have not heard of, operating in domains where the cost of a failure is real. The first paying customer converts because the operator-founder has existing credibility in the domain — they know the buyer, they speak the buyer's language, and they can describe the product's limitations honestly because they know the domain well enough to understand what they are. That combination of credibility and honesty is the thing an AI-capability-first team cannot fake, and it is the thing that converts a design partner conversation into a signed contract.

Sources

The Oracle modernisation moment: how AI is finally unlocking enterprise ERP

hello@knowmoreventures.com (Hareesh) — Sun, 15 Mar 2026 00:00:00 GMT

Seventy percent of Fortune 500 companies run Oracle. Most have never unlocked more than a fraction of what the system knows. Kanzi.ai is changing that — one support ticket at a time.

There is a specific kind of pain that every enterprise Oracle user knows. A procurement requisition is stuck. Nobody is sure why. The EBS workflow diagram has forty-two steps, the approval routing was last updated in 2019, and the one person who understood it left the company. The answer probably exists somewhere inside the system — in a configuration table, in an SOP document, in a Confluence page nobody has opened since 2021. Finding it used to take three hours and two phone calls. Kanzi.ai, a KnowMore Studio company, has built the AI system that finds it in ninety seconds.

Why Oracle is the right wedge

Oracle EBS is not a niche product. It runs the financial, procurement, HR, and supply-chain operations of a majority of the Global 2000. It is also, by almost any measure, the most under-exploited enterprise system in existence. The data is there: decades of transactions, approvals, rejections, and routing decisions. The knowledge is there: thousands of pages of MetaLink documents, SOP repositories, and consultant notes. What has been missing is an intelligent retrieval layer that can surface the right answer from the right source in plain English, at the moment someone needs it. That gap is Kanzi.ai's entire market.

Figure: A cross-section of a typical Fortune 500 Oracle EBS environment: eleven modules, four customisation layers, and a support queue that averages 340 open tickets. The AI assistant sits at the intersection, reading all three simultaneously.

The technical architecture is a multi-agent system built on Claude. A retrieval agent indexes the customer's internal Oracle documentation — MetaLink, SOPs, historical support resolutions, Slack threads — and serves as the knowledge layer. A synthesis agent takes the retrieval output and composes a structured answer, with citations. An escalation agent monitors confidence scores and routes low-confidence responses to a network of gig Oracle experts who verify and annotate the answer before it ships. The loop runs in under two minutes for most Tier-1 questions. Deflection rates in early deployments have crossed sixty percent within the first ninety days.

Most Oracle EBS problems are not engineering problems. They are knowledge-access problems. The system has the answer. Nobody could find it fast enough.
— Sai, Managing General Partner, KnowMore Ventures — portfolio review, March 2026

What this means for enterprise AI buyers

The department head — not the CIO — is now the primary buyer for AI tools that touch ERP workflow.
Deflection rate is the only KPI that matters in the first quarter. Enterprise buyers are not buying features; they are buying hours back.
Multi-agent systems that combine LLM retrieval with human expert escalation outperform pure-LLM products in regulated enterprise contexts.
The Oracle consultant network — 200,000+ globally — is an underexploited distribution channel for AI tools that speak the same language.

The opportunity extends well beyond support tickets. Once an enterprise has an AI system that understands its Oracle configuration at depth, the same substrate powers procurement workflow automation, period-close acceleration, and exception routing. Kanzi.ai's product roadmap follows that arc: start with the support ticket, earn the trust, expand into the workflow. The wedge is narrow by design. The adjacent market is enormous. For context on scale: a Fortune 500 Oracle estate typically generates between fifteen hundred and four thousand support interactions annually across finance, procurement, and supply chain. Deflecting sixty percent of those at a fraction of the cost of a helpdesk interaction produces a payback period measured in months, not years.

The human escalation layer is not a concession — it is the product. Kanzi.ai's gig expert network means that when the AI retrieval and synthesis agents produce a low-confidence answer, it does not surface an uncertain response to the customer. It routes to a credentialed Oracle consultant who answers within two hours and whose annotation becomes a training signal that improves future retrievals. This hybrid architecture is what allows Kanzi.ai to guarantee response SLAs that a pure-LLM product cannot. Enterprise procurement teams buy the SLA, not the model. Building the human layer first — before customers demanded it — is the decision that made the SLA promise credible at the first sales conversation.

Sources

India's $300B real estate market just found its AI moment

hello@knowmoreventures.com (Vishwa) — Tue, 10 Feb 2026 00:00:00 GMT

The Indian property transaction has always been opaque, slow, and paper-heavy. RealtyBlocks is building the AI stack that changes all three at once — and the timing has never been better.

India's residential real estate market transacts roughly $300 billion a year. It does so through a process that involves physical document verification, negotiation conducted entirely by phone, contracts drafted by hand, and valuations produced by individual brokers with no shared methodology. There is no Zillow. There is no Redfin. There is no shared MLS. What there is, increasingly, is a generation of urban buyers and sellers who are comfortable doing everything else on their phone — and who find the property transaction absurdly out of step. RealtyBlocks, a KnowMore Studio company, is the AI-native platform built for exactly that gap.

The four tools that matter

RealtyBlocks is not a listings portal. It is a transaction stack, and each of its four core products addresses a specific friction point in the Indian property cycle. PropGauge™ produces AI-driven valuations anchored in comparable transaction data, replacing the broker estimate that varies by thirty percent depending on whom you call. OfferChain™ digitises the negotiation, giving buyers and sellers a structured, time-stamped offer trail rather than a WhatsApp thread. AgreeMate™ generates compliant draft agreements, dramatically reducing the back-and-forth with a solicitor. SmartBid™ runs AI-coordinated auction processes for distressed or high-demand properties. Together, they convert a process that typically takes forty-five to ninety days into something that can be completed in under three weeks.

Figure: A stylised map of the Indian property transaction: seven intermediaries, fourteen documents, three government registries. The AI layer sits at each handoff, reading and writing simultaneously.

The structural tailwind is not just market size. The Indian government's push toward digital property registration — through RERA enforcement, digital stamp duty, and Aadhaar-linked verification — means the regulatory environment is actively pulling the transaction toward digital channels. Platforms that capture this shift early will have data advantages that compound quickly: every transaction closed is another training signal for valuation models, another pattern in the negotiation dataset, another precedent in the contract corpus.

The Indian property market is not waiting to be disrupted. It is waiting to be organised. The data exists. The buyers exist. What has been missing is an AI stack that speaks the market's language.
— Vishwa, GP AI Strategy, KnowMore Ventures — portfolio review, February 2026

What the PropTech AI opportunity looks like at scale

India's top eight cities collectively transact over one million residential units per year. Each transaction is a platform monetisation event.
RERA-compliant digital documentation is now mandatory in most states — a forcing function that pulls traditional brokers toward digital tools.
AI valuation models trained on closed transaction data compound in accuracy over time, widening the moat against manual comparables.
Cross-border demand — the Indian diaspora in the USA, UK, and Gulf — creates a natural India–USA corridor use case for remote transaction management.

The walkthesite.com platform, another KnowMore Studio build, extends the stack into the physical inspection layer — enabling AI-guided remote site visits that allow diaspora buyers and NRI investors to inspect a property from Dubai or San Jose. The two companies are complementary: RealtyBlocks handles the transaction; walkthesite handles the discovery and inspection. Together, they represent a full-stack AI real estate platform for the Indian market.

The broker network is not the competition — it is the distribution channel. Indian property transactions overwhelmingly involve a broker on at least one side. RealtyBlocks is designed to augment that broker, not replace them: PropGauge™ gives the broker a defensible valuation to anchor the conversation, OfferChain™ gives them a structured audit trail that protects them from disputes, and AgreeMate™ reduces the back-and-forth with solicitors that typically costs the broker three to four weeks of deal calendar. A broker who closes deals in three weeks instead of eight will close more than twice as many in a year. That is not a disruption pitch — it is a productivity pitch, and it converts broker skepticism into adoption faster than any alternative distribution strategy. The platform's early broker partnerships in Hyderabad and Bangalore confirm the pattern: adoption driven by referral within broker networks, without a single field sales hire.

Sources

When your enterprise workflow becomes an agent

hello@knowmoreventures.com (Kishore) — Tue, 20 Jan 2026 00:00:00 GMT

Multi-agent systems are no longer a research project. FlowAI is deploying Claude-based orchestrators across enterprise back offices — and compressing weeks of process time into minutes.

The phrase 'workflow automation' has been used so liberally — for everything from Zapier triggers to six-figure BPM implementations — that it has nearly lost its meaning. What FlowAI, a KnowMore Ventures portfolio company, is building is something categorically different: multi-agent systems that reason across enterprise processes, plan steps without being pre-scripted, and execute actions across multiple systems simultaneously. The difference between what FlowAI ships and what RPA vendors have sold for a decade is the difference between a script and a colleague.

What multi-agent enterprise automation actually looks like

A representative FlowAI deployment at a mid-market manufacturing company illustrates the pattern. The customer had a purchase order exception workflow that involved six humans across procurement, finance, and the vendor's accounts-receivable team, typically taking four to seven business days to resolve. The FlowAI orchestrator — built on Claude, with specialist agents for document parsing, ERP lookup, vendor communication, and exception classification — now handles seventy percent of exceptions end-to-end, without human intervention. The remaining thirty percent are escalated with a structured summary and a recommended action. Median resolution time fell from four days to six hours.

Figure: The agent stack: a thin orchestrator at the top, specialist agents below, a durable Postgres-backed queue running underneath. The human sits at the edge, receiving escalations rather than managing steps.

The architecture FlowAI uses is disciplined: a stateless orchestrator that owns policy (which agent runs when, and under what conditions), specialist agents with narrow responsibilities and no shared memory, and a durable queue that makes every step replayable. This separation is not an aesthetic choice. It is what makes production multi-agent systems debuggable. When something goes wrong — and it does — the team can replay the orchestrator's decision tape, identify the exact step that failed, and patch it without rearchitecting the system. The companies shipping multi-agent systems that break in production are almost always the ones where the orchestrator is also doing synthesis, classification, and retrieval simultaneously.

Every enterprise has workflows that are held together by email threads, shared spreadsheets, and institutional memory. That is not a technology problem. It is an AI opportunity.
— Kishore, GP Go-To-Market, KnowMore Ventures — FlowAI portfolio review, January 2026

Where the market is going

The addressable market is not RPA replacement — it is the long tail of workflows that were never automated because scripting them was too expensive.
Multi-agent systems outperform single-LLM tools on multi-step, multi-system workflows because they can reason about handoffs.
The eval harness — a library of recorded production traces with human-graded outcomes — is the primary moat. It cannot be reproduced from first principles.
Pilot-to-production conversion depends on deflection rate, not AI capability. Enterprise buyers measure hours saved, not model benchmarks.

The startup opportunity in multi-agent enterprise automation is genuinely large, and it is mostly unpenetrated. The Fortune 500 RPA market (UIPath, Automation Anywhere, Blue Prism) is worth roughly $12 billion and is built on brittle script-based systems that break whenever the underlying application changes its UI. That brittleness is the opening. Agent-based systems that navigate interfaces semantically — reading what is on the screen and deciding what to do next, rather than clicking pre-recorded coordinates — are inherently more robust. FlowAI's early enterprise deployments have not broken on a single UI update since go-live.

The pricing model also changes the sales conversation. RPA implementation projects were sold as capital expenditure: large upfront licences, multi-month implementations, and system-integrator fees that sometimes exceeded the software cost. FlowAI prices on workflow volume — a per-exception or per-automation-run basis — which converts the conversation from capex to opex, from IT to line-of-business, and from an eighteen-month procurement cycle to a ninety-day pilot with a clear success metric. That model shift is not just a pricing decision. It is a distribution strategy. The line-of-business buyer who can approve a monthly subscription without IT involvement is a fundamentally different and faster sales motion than the CIO who must justify a capital budget. And because the pricing scales with workflow volume, FlowAI's revenue grows automatically as customers expand usage — expansion revenue without an expansion sales motion.

Sources

Learning disguised as a game: how SparkLearn is fixing K-12 attention at scale

hello@knowmoreventures.com (Sharath) — Mon, 15 Dec 2025 00:00:00 GMT

The EdTech market is $400 billion. Most of it fails at the first problem: keeping a ten-year-old engaged for more than four minutes. SparkLearn, a KnowMore Studio company, solved that — and the implications for AI-native education are wide.

EdTech has a retention problem that its market size obscures. The category is enormous — roughly $400 billion globally, growing at fifteen percent a year — but the average student engagement rate on paid learning platforms sits under twenty minutes per week. Duolingo is the exception that proves the rule: it achieved genuine daily engagement by treating the lesson as a game level, not a lecture. The insight is not that gamification is a gimmick. The insight is that learning and play activate the same reward circuitry, and AI can now personalise both at scale. SparkLearn, a KnowMore Studio company, is the K-12 platform built on exactly that insight.

How SparkLearn works

SparkLearn's core product adapts to a child's grade level and national curriculum, then translates that curriculum into daily quests: short, structured learning sessions that feel like game levels rather than worksheets. Math becomes a dungeon to navigate. Reading comprehension becomes a mystery to solve. Science becomes a quest for resources. XP accrues. Badges unlock. The child's progress — not a static curriculum — determines what the AI surfaces next. The underlying model tracks not just right and wrong answers but the pattern of hesitation, retry, and skipping, adjusting difficulty in real time without making the adjustment visible to the child.

Figure: A child's learning session rendered as a quest map: three subjects, nine daily missions, one mascot character who narrates the journey. The AI adjusts the difficulty of each mission overnight, invisibly.

The business model follows the family rather than the district. SparkLearn sells directly to parents through a family account, bypassing the slow procurement cycles that have strangled most EdTech companies. That decision changes the feedback loop fundamentally. Parents report engagement weekly. Churn is visible within days. Product iterations ship based on parent and child behaviour rather than district IT schedules. The go-to-market is consumer, but the outcome — improved curriculum alignment and measurable grade-level progress — is institutional in quality.

The moment we stopped calling it studying and started calling it playing, the average session length doubled in one week.
— Sharath, GP Data & Analytics, KnowMore Ventures — SparkLearn portfolio review, December 2025

The AI-native EdTech opportunity

Adaptive curriculum — AI that adjusts to the child, rather than the child adapting to the curriculum — is the feature that incumbents cannot retrofit quickly.
Direct-to-parent distribution is ten times faster to iterate on than district procurement.
The India K-12 market — 250 million students, rising household income, mobile-first families — is the natural expansion market after US product-market fit.
Personalisation at the level of hesitation patterns (not just right/wrong) creates a dataset that compounds into a real moat.

The comparison that matters is not Duolingo. It is what Duolingo would be if it had a curriculum — if every lesson was tied to a specific grade-level standard, if the parent could see a report that mapped game progress to classroom readiness, and if the AI adjusted the game based on the child's school calendar. That is what SparkLearn is building. The gamification is not decoration. It is the mechanism by which a child chooses to open the app instead of YouTube, and that daily-active choice is the foundational product requirement from which everything else follows.

The data asset that accumulates underneath the game is the long-term moat. SparkLearn's AI tracks hesitation patterns, retry rates, time-on-task by subject and time-of-day, and the specific misconceptions that cause wrong answers — not to judge the child, but to predict the next ten minutes of content that will maximise retention. That dataset, labelled at the level of individual children's learning trajectories, is not replicable by a competitor who enters the market in two years. It requires two years of children playing the game to build. The flywheel is the retention mechanic: a child who plays every day generates more data, which makes the AI's predictions more accurate, which makes the game more engaging, which makes the child more likely to play every day. SparkLearn's earliest families have maintained daily active usage rates that the EdTech category has not historically sustained.

Sources

The golden hour of dating: how Saanjh is reimagining Indian matchmaking with AI

hello@knowmoreventures.com (Amala) — Fri, 14 Nov 2025 00:00:00 GMT

One curated introduction per evening at 6:14 PM. No swiping, no infinite scroll, no algorithmic anxiety. Saanjh is building the antidote to swipe fatigue — and proving that restraint is the sharpest product edge in consumer AI.

There is a moment every evening in Bombay when the light turns warm and the city exhales. It lasts perhaps twenty minutes. Everything that seemed urgent at noon feels negotiable. The founders of Saanjh named their product after that moment — saanjh, the Urdu word for dusk — and built the entire user experience around it. At 6:14 PM every day, each member receives one introduction. One person, chosen by the AI layer, surfaced at the exact moment the workday is ending. No queue of faces to scroll through. No gamified swipe mechanic rewarding speed over consideration. Just one name, one photograph, one opening line — and a window of time to decide whether to say hello.

The swipe-fatigue problem is structural, not cosmetic

The dominant metaphor of modern dating apps is the slot machine. Swipe left, swipe right, pull the lever, receive a variable reward. The mechanic was borrowed from casino game design — addictive in the short term and catastrophically bad for the thing it is supposed to produce. Dating app retention metrics and relationship formation metrics have always pointed in opposite directions. The apps that keep users engaged longest make matching feel abundant and frictionless; genuine connection is, by definition, the end of that engagement loop. The business model and the user outcome are structurally opposed.

Swipe fatigue is what happens when users have processed enough variable-reward loops to lose the ability to feel anything when a new face appears. Dating apps solved the discovery problem — there are always more profiles — and in doing so created an attention problem no feature polish has fixed. The founders of Saanjh started from a different premise: the problem is not that users lack options. The problem is too many options, delivered at the wrong moment, without enough context to make a meaningful choice.

We are not building a faster dating app. We are building a slower one. Restraint is the product. The 6:14 PM reveal is not a feature — it is the thesis.
— Amala, GP Design, KnowMore Ventures — Saanjh portfolio review, November 2025

The evening-reveal mechanic: design as moat

Saanjh's core interaction is deceptively simple. Every member completes an onboarding flow capturing not just demographics but values, pace-of-life orientation, and what they are actually looking for — not what they think they should be looking for. The AI curation layer surfaces one introduction per day at 6:14 PM local time. The moment is not arbitrary: user interviews consistently identified the end-of-workday transition as the emotional window when people felt most open and least defensive.

Members have until midnight to respond. If both express interest, a conversation opens. If either declines, the introduction closes — no rejection notification, no match-percentage display, no indication of whether the other person even opened the card. The founders spent two months debating mutual-match notifications and concluded that the social-anxiety cost of visible rejection was precisely what made swipe apps so exhausting. Saanjh removes that cost entirely.

Figure: The Saanjh evening reveal — one introduction at 6:14 PM, open until midnight. The interface is warm, unhurried, and deliberately sparse. Restraint as product philosophy.

The AI curation layer: why one match per day requires more intelligence, not less

The instinctive reaction to a one-match-per-day product is that it requires less AI sophistication. The opposite is true. When you give a user twenty options, algorithmic errors hide in the noise. When you give one option, every error is visible and costly. The matching quality bar is significantly higher.

Saanjh's curation model is built on a values-alignment layer above standard compatibility signals. Most dating algorithms optimise for engagement — swipes, messages, return sessions. Saanjh optimises for 'resonance likelihood': the probability that two people will find the introduction meaningful enough to continue the conversation. Resonance is a better proxy for user outcome and harder to game. You cannot fool a values-alignment model by uploading a better photograph.

The India dating market: size, structure, and why now

India's online dating market is projected to cross $280 million by 2030, growing at roughly 14% annually. The structural driver is the pace of change in how urban Indians approach partner selection. The arranged-marriage framework is being replaced, unevenly and sometimes painfully, by individual choice. The tools available — Western-designed swipe apps with superficial India localisation — are poorly matched to the emotional register of the people using them.

Saanjh is not targeting the mass market. The founding cohort is limited to 500 people across Bombay and Bengaluru — cities where product-market fit signal is clearest and cultural appetite for a considered alternative is highest. Of the 500 spots, 100 were claimed within the first four weeks of soft launch. The waiting list is substantially longer. Scarcity is part of the value proposition, and the team has shown discipline in not expanding faster than matching quality can sustain.

The operator thesis: why consumer social AI is a venture bet now

KnowMore's mandate has centred on enterprise and vertical AI — domains where operator-founder depth compounds into durable distribution advantages. Saanjh is our first Studio build in the consumer social category. The decision required us to articulate why consumer social AI is a different bet in 2025 than it was in 2018.

The answer is a capability threshold. For most of the last decade, consumer social apps chose between human curation — expensive, slow, unscalable — and algorithmic recommendation — cheap, fast, optimised for engagement over outcome. The AI layer available in 2025 enables a third path: personalised, values-aware curation that improves with every data point, delivered at algorithmic cost. Saanjh is built on that shift. The founders are not using AI to make swiping faster. They are using AI to make genuine curation viable at consumer price points.

The founding-member model is a thesis on consumer social moats. The apps that defined the last decade — Tinder, Bumble, Hinge — won on network effects. Building a competing network effect from scratch is nearly impossible. What is possible is a reputation effect: a product distinctive enough that users talk about it unprompted, and exclusive enough that demand exceeds supply. Saanjh's 500-person cohort is a moat-building exercise dressed as a product launch. The waiting list longer than the cohort is not a scaling failure. It is evidence the reputation effect is working.

Sources

Property search is broken. AI is fixing it from the ground up.

hello@knowmoreventures.com (Amala) — Sat, 08 Nov 2025 00:00:00 GMT

India's real estate discovery problem is not a data problem. It is an intelligence problem. realestateindia.ai is building the search layer the market was missing — and the window for incumbents to respond is closing.

There are roughly forty million property listings in India at any given time. They are distributed across MagicBricks, 99acres, Housing.com, hundreds of builder microsites, and thousands of broker WhatsApp groups. They are described in inconsistent formats, priced without comparables, photographed by whoever had a phone that day, and tagged with area labels that mean different things on different platforms. Finding the right property — for a specific buyer profile, a specific price band, a specific commute radius — requires either a patient broker or a very tolerant Sunday afternoon. realestateindia.ai is building the AI search layer that makes neither of those necessary.

What AI search means in practice

The core product aggregates listings from across India's fragmented property data sources, normalises them into a unified schema, and layers an intelligent search interface on top that understands natural language queries at the level of buyer intent. 'A two-bedroom flat within thirty minutes of Hitech City, under seventy lakhs, in a society with a gym and covered parking' returns a ranked, filtered result set in seconds — not a generic list that the buyer then manually filters through seven dropdowns. The ranking model factors in price trends, builder reputation, proximity to verified amenities, and recency of listing activity. The gap between a query and a qualified result collapses.

Figure: The data pipeline: forty million listings, five aggregation sources, one unified schema, and an intent-aware ranking model sitting at the top. The buyer types a sentence and gets a shortlist.

The strategic insight is that the listings portals have the data and have largely exhausted what keyword search and manual filters can do with it. The AI layer does not replace the portals — it reads them, interprets them, and serves buyers who have already learned not to trust what a raw listing says. Market-level intelligence compounds quickly. A model that has seen ten million searches, their refinements, and the properties buyers ultimately contacted has a very detailed picture of the gap between what listings claim and what buyers actually want. That gap is the core product.

Buyers in India are not looking for listings. They are looking for reassurance. They want to know that the price is real, the builder will deliver, and the commute is actually what the listing says it is. AI can provide that reassurance at scale.
— Amala, GP Design, KnowMore Ventures — realestateindia.ai portfolio review, November 2025

The opportunity for AI-native property search

Natural language search is not a UX preference — it is a fundamental product requirement for a market where buyers do not know the right filter keywords.
Aggregation plus intelligence is a defensible position: the portal cannot replicate the intelligence without the aggregation data, and the data is hard to aggregate.
Builder reputation scoring — trained on delivery timelines, RERA complaints, and buyer reviews — is the feature that converts search to qualified inquiry.
Mobile-first architecture is non-negotiable in the Indian market. Sixty percent of property searches happen on a phone.

The parallel to what Zillow did to the US market is instructive but imperfect. Zillow succeeded by aggregating MLS data that was already structured. The Indian market has no MLS — it has WhatsApp groups, broker networks, and portals with inconsistent schemas. The aggregation problem is an order of magnitude harder, which is precisely why no incumbent has solved it well. That difficulty is the moat: any competitor who wants to replicate the intelligence layer has to solve the data problem first, and the data problem takes years. realestateindia.ai has a head start of exactly that kind.

The monetisation model also benefits from the intelligence layer. Listings portals earn revenue from featured placement — a model that rewards the highest bidder, not the most relevant property. realestateindia.ai's ranking algorithm surfaces the most qualified matches for a specific buyer query, which means that the most relevant listing earns the top slot without paying for it. That changes the value proposition for brokers and builders who use the platform: they are not buying eyeballs, they are buying qualified inquiry from buyers who match their property profile. Conversion rates on qualified inquiries are measurably higher than on broad impressions, and higher conversion rates are the metric that justifies the platform's fee. The intelligence layer is the product that enables the business model, not a feature added on top of it.

Sources

Exploring homes like a game: how DeReal is turning property discovery into an immersive AI experience

hello@knowmoreventures.com (Vishnu) — Wed, 15 Oct 2025 00:00:00 GMT

Property search is one of the most consequential decisions a buyer makes, yet the experience of searching ranks among the worst in consumer technology. DeReal — in beta — is the first platform to treat property discovery as an attention problem, and to solve it.

The global residential property portal is one of the most visited categories on the internet and one of the least engaging. A typical buyer session involves loading a map, adjusting three price sliders, scrolling through forty listings that look identical in their thumbnail photography, clicking four of them, reading the same square-footage and bedroom count formatted in four different ways, and closing the tab. The session produces almost no useful information and no emotional conviction about any property. Real decisions happen later, at a physical site visit, when the buyer finally has enough sensory data to form a view. DeReal, a KnowMore Ventures portfolio company currently in beta, starts from the premise that this sequence is backwards — and that AI can collapse the gap between first search and genuine conviction without a single in-person visit.

How DeReal works

DeReal turns the property search session into a structured, immersive experience with game-design principles applied at every layer. On entry, the platform's AI matching engine builds a preference profile from the buyer's initial inputs — budget, location range, lifestyle signals — and surfaces a ranked shortlist of properties tailored to that profile, not a raw feed filtered by price. Each property is presented as an explorable environment rather than a static listing: spatial photography, annotated room-by-room breakdowns, AI-generated neighbourhood commentary, and a live consultation option that connects the buyer with a professional in real time over video. The session is structured as a progression. The buyer does not scroll; they explore.

Figure: The DeReal player map: five roles arranged around a single property — buyer, seller, buyer's agent, listing agent, and video consultant — each with their own view of the same transaction, coordinated by the platform's AI layer.

The five-player-role architecture is DeReal's most distinctive structural decision. The platform is not a buyer-facing portal with an agent bolt-on. It is a coordination layer for every participant in a property transaction: the buyer exploring listings, the seller monitoring interest through an owner lead-pipeline dashboard, the buyer's agent advising in real time via the integrated video consultation channel, the listing agent managing their property's presentation, and the professional consultant (mortgage advisor, inspector, surveyor) available on demand through the platform's live video module. Each role has its own view of the same property and transaction — the buyer sees a discovery experience, the seller sees a qualified pipeline, the agent sees an engaged client. The AI summarises every session and posts a structured brief to the relevant parties, so no context is lost between sessions.

The zero-account-required entry point is a deliberate beta-stage choice. Most property portals require registration before revealing full listing details — a friction point that eliminates a significant proportion of early-session buyers who are not yet ready to identify themselves. DeReal allows full exploration without a login, then captures identity at the point of genuine intent: when the buyer requests a live consultation or saves a property to their shortlist. This sequence mirrors the conversion logic of successful consumer apps: demonstrate value before asking for commitment. In beta, the platform is measuring the gap between anonymous exploration sessions and identified-buyer conversion events, using that data to calibrate how much friction to introduce and when.

Property portals have solved the inventory problem. Nobody has solved the engagement problem. DeReal is the first platform we have seen that treats a buyer's attention as the scarce resource it actually is.
— Vishnu, GP Partnerships, KnowMore Ventures — DeReal portfolio review, October 2025

What the beta is testing and why it matters

Session depth is the primary beta metric: how many properties does a buyer explore, and how long does each exploration last before intent signals appear.
Live video consultation conversion is the secondary metric: the share of explorers who request a real-time professional, which is the platform's core revenue event.
Owner lead-pipeline engagement measures whether sellers actively monitor their listing's activity — a signal of whether the platform creates supply-side retention.
AI summary quality is being evaluated against buyer recall: do buyers remember, two days later, what they learned from a DeReal session that they would not have learned from a standard portal?

The competitive framing for DeReal is not the property portal — it is the attention economy. The platform competes less against Zillow or MagicBricks than against any digital experience that bids for a buyer's ninety-minute Saturday morning slot. A buyer who spends that slot on DeReal instead of Instagram or a property portal has given the platform something far more valuable than a click: a sustained, high-intent engagement session that the AI can read, annotate, and act on. The Comms Centre that centralises all buyer-agent-seller communication inside the platform is the retention mechanic that makes that attention compound: buyers return to check responses, sellers return to see pipeline movement, agents return to manage multiple clients. The session is the product. The transaction is the outcome.

The India and diaspora buyer segments are DeReal's most acute early use cases. For an NRI buyer evaluating a property in Hyderabad from a San Jose apartment, the choice has historically been between an unreliable listing, a video call with a broker of uncertain trust, and a physical visit that costs a return flight. DeReal's combination of spatially rich exploration, AI-generated due-diligence commentary, and on-demand video consultation with a verified local professional collapses that choice into a single session. Beta feedback from this segment shows the highest consultation-conversion rates and the longest average session lengths — two signals that suggest the platform's value proposition is strongest precisely where the information asymmetry is highest.

Sources

The memory layer: why the next AI moat is what your system remembers

hello@knowmoreventures.com (Purush) — Thu, 02 Oct 2025 00:00:00 GMT

Language models are amnesiac by default. memtra.ai is building the persistent knowledge graph that makes AI assistants genuinely contextual — and the market for that capability is larger than most people assume.

Every conversation with a language model begins from zero. The model does not know who you are, what you worked on yesterday, what your organisation's vocabulary means, or what decision you made last week that makes today's question consequential. The context window is large but it is also ephemeral. The moment the session ends, everything is gone. For consumer use cases — a quick summary, a draft email — this amnesiac default is tolerable. For enterprise use cases — a weekly planning assistant, a long-running procurement analyst, a personal research companion — it is the difference between a tool and a toy. memtra.ai, a KnowMore Studio company, is building the memory layer that bridges that gap.

What a knowledge graph adds to a language model

memtra.ai's architecture is a personal and enterprise knowledge graph that persists context across sessions, across users, and across applications. The graph stores entities (people, projects, concepts, decisions), their relationships, and the temporal context that makes relationships meaningful. When a user asks their AI assistant a question next Tuesday that relates to something they discussed last Thursday, the memory layer surfaces the relevant context automatically — not by storing a conversation transcript (which creates noise), but by extracting the structured entities and relationships that matter and making them retrievable in milliseconds. The model receives a query plus a context injection that would otherwise take the user several minutes to reconstruct manually.

Figure: A knowledge graph node cluster: a user's three active projects, seven relevant relationships, and four pending decisions, rendered as a constellation. The AI navigates this map before it answers.

The enterprise application is especially clear. An enterprise knowledge graph that remembers which vendor relationships are active, which contract negotiations are in progress, which approvals are pending, and which team members own which decisions transforms a general-purpose AI assistant into a genuinely contextual colleague. The same assistant that today requires the user to re-explain their situation every session becomes, with the memory layer, an assistant that already knows what the user means when they say 'the Mumbai deal' or 'the Q3 headcount ask.' That contextual shorthand is worth far more than any improvement in the underlying model's reasoning capability.

Memory is not a feature. It is the fundamental requirement for AI to be genuinely useful in a professional context. Without it, every session is a first meeting.
— Purush, GP Product, KnowMore Ventures — memtra.ai portfolio review, October 2025

The market for persistent AI context

Enterprise knowledge graphs are a greenfield market — the incumbents (Notion, Confluence, SharePoint) store documents, not relationships.
Personal AI assistants with memory will command a premium because the switching cost compounds with every session of context they accumulate.
The API opportunity is significant: every AI application that wants to offer persistent context is a potential customer for an off-the-shelf memory layer.
Privacy-preserving graph design — storing relationships without storing raw conversation content — resolves most enterprise data-handling objections.

The near-term competitive risk is that frontier model providers — OpenAI, Anthropic, Google — add memory features natively to their products. That risk is real but bounded: native model memory is optimised for general-purpose recall, not for the structured entity-relationship graphs that enterprise contexts require. The difference between 'I remember we talked about the Mumbai deal' and 'I can retrieve the current state of the Mumbai deal's contract negotiation, the three open items, the counterparty relationship history, and the approval pending from the CFO' is a product gap that a model memory feature does not close. The structured graph is the product, not the recall.

The India angle is specific and underappreciated. Indian enterprises — mid-market manufacturing companies, professional services firms, family conglomerates — operate with institutional knowledge that lives almost entirely in individuals' heads rather than in documented systems. When a senior relationship manager or a procurement head leaves, the organisation loses not just their time but their entire context library: which vendor negotiations are in progress, which customers have implicit credit arrangements, which supply chain decisions are pending. memtra.ai's knowledge graph is the institutional memory layer that these organisations have never had access to before. The product is not a general-purpose AI assistant with memory. It is the organisational context database that prevents institutional amnesia — and in the Indian mid-market, where documentation practices lag the volume of transactions, that is an acute pain that founders are willing to pay to solve.

Sources

KYC's AI upgrade: why property verification in India is about to get dramatically faster

hello@knowmoreventures.com (Srinivas) — Tue, 22 Jul 2025 00:00:00 GMT

Identity and property verification is the number-one friction point in Indian real estate transactions. ClearPass is the AI layer that removes it — and the regulatory environment is pulling directly in its direction.

The average Indian residential property transaction involves verifying the identity of at least two parties, confirming the chain of title across sometimes decades of ownership transfers, validating encumbrance certificates from the relevant sub-registrar, and cross-referencing property records against government databases that are not always current, consistent, or digitised. This verification process, done manually, takes between seven and twenty-one days depending on the state and the complexity of the title chain. It is also the step most likely to surface a problem that kills a deal — a missing link in the chain, a disputed boundary, an encumbrance that the seller did not disclose. ClearPass, a KnowMore Studio company, is the AI platform that automates this entire verification layer.

How ClearPass works

ClearPass's platform ingests identity documents (Aadhaar, PAN, passport), property records (sale deeds, encumbrance certificates, khata extracts, RERA registration documents), and cross-references them against a network of government data sources using AI-assisted document parsing, OCR, and entity resolution. The platform identifies discrepancies — mismatched names across documents, unresolved encumbrances, expired NOCs — and surfaces them as structured exceptions rather than burying them in a stack of raw documents. A KYC check that would have taken a paralegal three days now completes in under four hours. For RERA-compliant transactions, the structured output is already in the format regulators require.

Figure: A typical ClearPass verification run: fourteen input documents, six government database lookups, three discrepancies surfaced automatically, one human review required. Total elapsed time: three hours and forty minutes.

The regulatory tailwind is significant. India's Digital Personal Data Protection Act (2023) and the ongoing push toward Aadhaar-linked property registration are both pulling the market toward structured digital verification. State governments are progressively digitising sub-registrar records; RERA mandates are requiring developers to maintain verified documentation for all projects. These regulatory developments do not create the ClearPass opportunity — the manual verification problem existed long before any regulation — but they accelerate the market's willingness to pay for an AI-driven solution. Compliance is becoming a buying trigger.

The single most common reason a property deal collapses in India is a verification problem that was discoverable in the first week but found in the last. AI changes that timeline entirely.
— Srinivas, GP Operations, KnowMore Ventures — ClearPass portfolio review, July 2025

The KYC AI opportunity in Indian real estate

India processes approximately 1.2 million property registrations per month. Each one is a potential ClearPass transaction.
The cost of a failed transaction — legal fees, stamp duty on re-registration, broker commissions — can exceed two percent of the property value. Verification automation pays for itself on the first avoided failure.
Banks and NBFCs processing mortgage applications face the same verification bottleneck. Lender integrations are a natural expansion channel.
API-first architecture — ClearPass as a verification service for banks, builders, and brokers — opens the market beyond direct-to-consumer.

The cross-border dimension is also real. NRI buyers purchasing property in India from the USA, UK, or Gulf face a verification process that is nearly impossible to complete remotely without trusted local intermediaries. ClearPass's digital verification layer enables a diaspora buyer in San Jose to complete the identity and property check on their phone without flying to Chennai for a document review. Combined with the RealtyBlocks transaction stack and walkthesite's remote inspection capability, the result is a full-stack remote property transaction that was not possible eighteen months ago.

State-level variation in government database quality is the core technical challenge ClearPass has had to solve, and solving it is the barrier that protects the platform from fast-follower replication. Karnataka, Maharashtra, and Telangana have relatively mature digital sub-registrar records; Bihar, Uttar Pradesh, and parts of Tamil Nadu have partially digitised registries with inconsistent schemas. ClearPass's entity resolution layer normalises across these variations automatically, mapping document fields to a unified property schema regardless of the originating state registry's format. Building that normalisation layer required eighteen months of data collection across twenty-two state systems. A competitor starting today faces the same eighteen months of work, plus the regulatory relationships with state land departments that ClearPass has already established. The moat is not the AI. It is the data infrastructure the AI runs on.

Sources

Walking a site without leaving your desk: how virtual tours are becoming AI copilots

hello@knowmoreventures.com (Vishnu) — Tue, 10 Jun 2025 00:00:00 GMT

The property site visit is one of the most expensive, time-consuming steps in real estate. walkthesite.com is turning it into an on-demand AI experience — and opening Indian property to global buyers.

The property site visit is a remarkably inefficient event. The buyer travels to the property at a time that suits the broker, walks through rooms that may or may not be furnished, asks questions from memory rather than from preparation, and leaves with a set of impressions that are difficult to share with a spouse, a parent, or a financial advisor who was not present. The visit is also geographically constraining: a buyer in London cannot visit a flat in Hyderabad without a flight. The technology for remote property inspection has existed in various forms for a decade — 360-degree cameras, video walkthroughs, Matterport scans — but the experience has never been intelligent. You could look at the space remotely. You could not ask it questions. walkthesite.com is changing that.

What AI-guided virtual inspection looks like

walkthesite.com combines 3D spatial capture with an AI guide that narrates the property from the buyer's perspective. The visitor moves through the virtual space and can ask questions in natural language: 'What is the natural light like in the living room at 4 pm?' 'Is the kitchen ventilation sufficient for Indian cooking?' 'What is the construction quality rating on this finish?' 'Is this floor plan typical for this developer?' The AI guide draws on the property's spatial data, the builder's specification sheet, comparable properties in the same development, and public RERA data to produce answers that a standard video walkthrough cannot. The result is a site visit that the buyer can conduct in twenty minutes, at any time, from any location, and share a link of with anyone.

Figure: A first-person view of a virtual site walk: the AI guide's dialogue overlaid on the 3D model of a two-bedroom flat in Pune. The buyer has asked about the kitchen window's alignment. The answer is already on screen.

The NRI and diaspora buyer is the most acute use case, and it illustrates the broader opportunity. An Indian-American family in the San Francisco Bay Area looking to purchase a property in Bangalore for investment or for a parent's retirement is currently constrained to either flying to India, trusting a local relative's judgment, or buying based on a video call with a broker. walkthesite's virtual inspection layer removes all three constraints. The family can conduct a thorough inspection at 9 pm Pacific Standard Time, involve all decision-makers simultaneously, and generate a structured report to share with their accountant and legal advisor. The property transaction's final geographic friction point dissolves.

We are not building a prettier video tour. We are building the AI property advisor that answers every question a buyer would have asked in person — and several they would not have thought to ask.
— Vishnu, GP Partnerships, KnowMore Ventures — walkthesite portfolio review, June 2025

The virtual inspection market and its AI moment

The Indian diaspora in the USA alone comprises 4.5 million people, with significant property investment in India. That is a large, underserved buyer segment for remote transaction tools.
AI-guided inspection — not just spatial capture — is the feature that converts a virtual tour from a media format into a decision-making tool.
Builder adoption is a B2B distribution channel: developers who offer AI-guided virtual tours on pre-launch projects close faster and with fewer costly in-person visits.
The combination of virtual inspection (walkthesite), AI search (realestateindia.ai), and transaction automation (RealtyBlocks) represents a full-stack remote property purchase for the first time in the Indian market.

The infrastructure requirements have also changed in walkthesite's favour. 3D spatial capture hardware has become commodity-grade — a modern smartphone with a LiDAR sensor produces data quality that would have required a $20,000 Matterport camera four years ago. AI inference costs have fallen to the point where a natural-language question answered by a property-specific model costs fractions of a cent. The combination of cheap spatial capture and cheap intelligent inference is the inflection point the virtual inspection market has been waiting for since the first 360-degree camera tour launched a decade ago.