Beyond Vibe Coding: Adam Rian - AKTUS AI

<p>Adam is the founder and CEO of AKTUS AI, a Silicon Valley company building custom AI systems for finance and real estate — and a firm Trion has been working with directly. He has a PhD in computer science, spent 15 years in AI and machine learning, and was product lead for data and ML at Twitter before co-founding Flowcode.<br>In this episode, David and Adam get into why out-of-the-box AI keeps failing on real estate work, what actually separates a prototype from a production system, and why Adam thinks the org chart itself is the thing AI ends up rewriting.<br><strong>What we get into</strong><br><br><strong>Why generic AI can&#39;t do real estate.</strong> The models were trained on public data. Real estate runs on proprietary data, domain judgment, and a secret sauce that was never on the internet — and it&#39;s multidisciplinary on top of that: finance, legal, architectural, relationship-driven.<br><strong>Vibe coding: 100% and 0%.</strong> Adam&#39;s line is blunt. Vibe coding is 100% the right tool for prototyping and requirement gathering — it replaced weeks of whiteboarding and Figma mockups. For production code, it&#39;s 0%. None of it survives. He explains why, and it comes down to edge cases and maintainability.<br><strong>The maintenance trap.</strong> David&#39;s own experience: building something that works, then spending half your day keeping it alive instead of doing the job. The productivity gain never shows up.<br><strong>Digital teammates, not software.</strong> You&#39;re not buying an application. You&#39;re onboarding something you train, give feedback to, and measure — the same way you&#39;d manage a first-year analyst.<br><strong>How you actually measure success.</strong> Booking a meeting is not the same as booking the right meeting. Adam walks through how AKTUS defines KPIs with a client during onboarding, and why measurement has to be designed in from day one rather than bolted on.<br><strong>Why AI hallucinates, explained simply.</strong> Context windows, the myth of the million-token window (only about 10% of it is effective), and why AI forgets the middle of a document the same way you forget the middle of a novel.<br><strong>Context engineering — and hiding data on purpose.</strong> Sometimes the right move is keeping data <em>away</em> from the model. Adam explains the system AKTUS built to ingest thousands of leases, zoning docs, and a 60-tab underwriting model without the whole thing falling apart.<br><strong>Five years