START

Liam Karlsson had no clue why his SEO clients were losing traffic while rankings heldThen his 57-year-old mom asked ChatGPT for new running shoesNike answer. Bought the shoe. Was super happy. Google was never part of that customer journey...That was the seed of Rankad.aiHe pitched Co-Founder William Gyltman. They went all inTrack and grow brand visibility across ChatGPT, Perplexity, Gemini, and Copilot. Enter your domain, 30 seconds later you're in the app. AI agent optimizes your site directlyLiam Karlsson & William Gyltman, Co-Founders, Rankad.ai at The Residency Demo Day 🎙️ Fondo START pod w/ David J. Phillips (full ep in comments)00:25 “We’re helping companies earn more money in AI search” 02:34 AI search is moving fast: new models, algorithms, protocols 03:05 A new source of income beyond Google traffic 03:31 AI search is growing fast, but has not outrun Google yet 05:34 Enter your domain and get inside the app in about 30 seconds 05:52 Visibility, competitor, and company data inside one platform 06:07 AI gives optimization tasks based on scan data 06:15 The agent executes tasks on your site 09:41 SEO rankings held, but traffic still disappeared 09:47 Liam’s mom used ChatGPT to buy running shoes 10:00 “That was like the seed of Rankad”

Amazon sellers do not have a data problem.They have a decision problem.They already have the clicksThe conversionsThe impressionsThe keyword historyWhat they do not have is a system that knows what to do with it.So brands pay agencies $5,000 to $50,000 a month - and still lose money on adsMatthew Chen built Laurence to change that.A quantitative system for Amazon advertising.Built for continuous decision-making under profit constraints.Using existing ad copy, reinforcement learning, and custom models to run ads on autopilot.When confidence is high, it actsWhen data is sparse, it borrows signal from similar keywordsThe result: about 40% better performance for customers.Amazon is the wedge.Autonomous performance marketing is the bigger vision.🎧 START pod: Matthew Chen, Founder & CEO, Laurence "Autonomous performance marketing"00:23 What Laurence does02:12 How the company found the wedge02:47 The customer quote that changed the company03:15 The flaw in the standard Amazon ad model05:25 The move from Amazon to the wider internet06:12 How Laurence uses transformers today

Construction doesn't fail on the jobsite. It fails in the drawings. The jobsite just reveals itProject managers spend half their time scanning plans page by page for conflicts between disciplines. Plumbing through steel beams. Electrical into HVACThey still miss most of it. Millions in rework when caught in the fieldMilind Sagaram built Articulate to catch these issues before construction startsAI reads the PDFs. Finds clashes across architectural, structural, and MEP sheets. Generates draft issue reports automaticallyThe surprise: construction teams aren't resistant. They want it more than anyone expected🎙️ Milind Sagaram, Co-Founder & CEO, Articulate / Helonic.com on Fondo START pod 00:18 AI for finding drawing issues before construction starts02:24 The old workflow: manual plan review02:45 The consequence: rework and delays03:00 Small issues, massive downstream cost03:27 Copilot, not replacement04:04 Why AI belongs in construction04:39 What surprised him about selling into the industry05:59 Who Articulate sells to

Agents don’t need bigger models. They need better tools.Morph trains coding subagents.Not for humans. For frontier models.Fast Apply edits at 10,000 tokens/sec. WarpGrep handles code and log search.Both keep the main model’s context cleanBecause when context gets too large, performance drops.Now Morph is pushing coding subagents even faster.One newer model runs at 33,000 tokens/sec: https://docs.morphllm.com/sdk/components/compact🎙️ Tejas Bhakta, Founder & CEO, Morph01:30 Fast Apply + WarpGrep02:26 Context fills up around 100k02:38 Keep the main model context clean03:31 “You can’t scale human attention 100x”07:28 Founders are missing how to value human attention08:30 New model at 33,000 tokens/sec10:08 Better, faster, and cheaper than the frontier

Pamir Ehsas spent years as outside counsel serving startups. He saw the same problem on repeatSimple legal work took weeks. Pricing was opaque. Lawyers kept starting from scratch instead of using AISo he built Arcline. AI generates the first draft. Elite lawyers from the best firms, and schools (Harvard, Oxford etc.) do the final revisionUp to 80% of the work gone. Same-day turnaround (Try getting that from a traditional law firm) 50+ venture-backed startups already onboard when this ep was recordedMost legal AI tries to replace the lawyer. Pamir replaced the busywork02:09 “Fondo, but for legal” / AI-native legal for startups03:12 Sold AI to law firms first; then pivoted to end users04:52 Why Arcline uses experienced lawyers, not junior lawyers05:43 Faster, higher-quality work — including complex matters07:00 Big Law was slow; incentives were broken07:36 Why going direct creates the data loop to improve AI08:34 AI does the grunt work; lawyers create trust11:08 What startups actually come to Arcline for12:26 The long-term vision: legal advice at your fingertips

An employee used his corporate Gemini account to generate fake receipts for reimbursementsNot because he was lying. He just didn't have the real onesThat's when the CISO realized they needed OximyAsk an enterprise how many AI tools they use. They say 10. The real number is probably 40+Most of the risk isn't malicious. People just want to get things done faster. So customer lists end up on free tools with no DPAThe first instinct is to block everything. But people bypass restrictions anywayThe real question: how do you say yes to AI without losing control?That's Oximy. A control layer for enterprise AI adoption. Track usage, manage spend, enforce governance🎙️ Naman Ambavi, Founder & CEO, Oximy on Fondo START pod00:19 — What Oximy helps enterprises understand and manage02:24 — From India to the Bay, and the thesis behind Oximy03:29 — Track adoption, cost, and risk controls03:52 — “They say they’re using not more than 10… the number goes roughly over 40”04:22 — Compliance, retraining risk, and why oversight matters05:10 — Most misuse starts with harmless intent06:23 — The gap between Silicon Valley and corporate America07:35 — How to say yes to AI without losing control10:25 — Why the pain shows up most clearly at 1,000+ employees12:59 — Why banning AI is the first instinct — and why it doesn’t last

Attorneys are trained to spot issues. That’s literally what law school teaches.Show them your product, and the first thing they’ll say is: “the margin is off on this.”Every hour they spend learning software is an hour they’re not billing.Raffi Isanians knows that because he lived it.Kirkland. Gunderson. Years inside private equity and venture work.That’s why Mage Legal has a simple standard: if a lawyer opens the product with no instructions and can’t figure it out, "we’re failing"Comprehensive AI coverage across the entire data room:red and yellow flags, diligence memos, disclosure schedules, redline comparison. 1,500 documents in tens of minutes. All async.05:15 - Puts a TOS into ChatGPT 3.5, comes back 85-90% there06:20 - Every hour learning a tool is an hour not billed08:37 - The product worked, nobody wanted it09:00 - YC partner on the two-year window09:53 - AI is an F1 engine in a world of bicycles13:00 - Clients are pushing AI adoption, not the lawyers15:19 - The goal is zero behavior change17:00 - 1,500 documents, diligence memos, tens of minutes18:22 - One good associate now does the work of six30:56 - Engineers simplify, lawyers complicate33:00 - 11PM, picture of his daughter, back to work40:34 - Simple enough to use with zero instructions

The internet was the training set for intelligenceNobody has built the equivalent for the physical worldPreviously, Lucas Ngoo co-founded Carousell, scaled it past $1B Now at Cortex AI he's collecting the data robotics labs need to train foundation modelsCameras, VR headsets, glasses on factory workers, retail workers, everyday people. Recording real-world manipulation work (Maybe tens of millions, even hundreds of millions of hours)Not building the robot. Not training the model. Collecting what goes in2026: scale data in a big way 2027+: start rolling robots out, 90% teleoperation, 10% autonomyWhile humans step in, the system keeps learningThat's the flywheel toward full autonomy01:38 Cortex as the data layer for general-purpose robotics02:19 Why text data doesn’t transfer cleanly to the real world02:51 Recording day-to-day human manipulation work03:23 “Tens of millions” to “hundreds of millions of hours” of data06:37 Cortex’s role: “We are really the data piece”07:28 When humanoids become part of everyday life08:00 2026 as the year data scales08:13 90% teleoperation, 10% autonomy08:37 Data flywheel toward full autonomy08:55 Synthetic data + real-world data reinforcing each other

Some hotel software is still DOS-based. Sometimes pen and paper (that's why guests are waiting 45 mins to get towels)Gavin's dad has worked at Marriott for the last nine years. When he showed Gavin the old software, that was the sparkLance builds AI agents that answer calls, handle back office operations, run sales workflows & moreStarted with voice, got inside the hotels, and realized how much more they could automateWith coding agents one engineer acts like fiveBig contracts need custom solutions. Now they can deliver🎙️ Gavin Brennen, Co-Founder, Lance (YC W26) on Fondo START pod01:02 The reality of hotel software01:53 The insight that sparked the company02:01 A better way to handle guest requests02:46 Why Lance had to build desktop agents03:08 The centralized dashboard layer04:00 Why hotel problems are highly custom04:32 One engineer can now act like five05:51 “Go all in”

Direct sales reply rates are going down. AI spam is making it worseNikhil Reddy saw the shift early: as trust matters more, partnerships become a real revenue channelProblem is most partnership teams are still running on spreadsheets. They don't know where to focus. Attribution across emails, events, and co-marketing is a messArzule uses CRM data, market signals, and ecosystem signals to help partnership teams discover and prioritize the partnerships that actually drive revenueTwo people building it. Already working with companies generating over $400M in ARRApplied late to YC's fall batch, never even got a reply. Applied again for winter, got in, and their group partner told them to pivot the next day. That became Arzule🎙️ Nikhil Reddy, Co-Founder & CEO, Arzule on the Fondo START pod01:09 What Arzule does and who it's for02:24 Direct sales is dying. Trust is taking over02:43 Pivoting mid-batch from multi-agent coordination to partnerships03:12 Why partnership teams are finally getting their moment03:54 The data layer: CRM signals, ecosystem signals, company-specific inputs04:09 Why attributing revenue to partnerships is so hard05:24 Affiliate links, rev-share, bounties, and contracts in one platform06:03 What onboarding looks like for larger customers07:02 Start analyzing everything early, even if it feels relationship-driven07:48 Tracking signals like new partnership hires and market expansion08:28 Revenue attribution as the core of predictable partnerships09:26 Applied late, got no reply, applied again, pivoted the day after getting in10:02 Move fast and pick something you want to do for 10 years

Setting up a single trigger in Zendesk takes 30 clicksWith Pillar it takes one sentenceJJ Maxwell built an open source copilot you build into your app. Users talk to it in natural language and it drives the app for themThe problem: products can do a lot but users don't always know what's there. So they ask support. Or they churnBefore Pillar, JJ built a creator ad marketplace with about 40,000 creators. Then spent two years on another product through YC W24. About $30M on the platform, real users, decent growth. Pivoted anyway. As soon as he lost belief it was gonna work, he ripped the bandaid offWeb MCP is already rolling out. Companies trying to stop agents from taking actions are fighting a losing battle🎙️ JJ Maxwell, Founder & CEO of Pillar (trypillar.com) on the Fondo START pod01:56 What Pillar is: a copilot that’s easy to build into your app02:19 Why users ask support or churn when product complexity hides value03:05 “30 clicks” in Zendesk becomes a sentence03:39 Why AI can do this now: models are better at reasoning and chaining actions04:19 What implementation looks like: wrap existing frontend code and tool calls05:25 Why teams can often get Pillar working in about a day06:09 The Double journey, real traction, and the decision to pivot11:00 Web MCP and why agent-ready software is coming12:23 Why companies may not be able to stop agents from taking actions forever