TwoSetAI interview. Sam Kececi, founder and CEO of The Sentience Company, talks with host Angelina Yang about building a personal AI that doesn’t just write like you but acts for you — answering your team, sitting in on interviews, closing deals. Around 50 minutes; this is the founder-level companion to the Workshop #5 engineering talk from his colleague Ahmet İlten.

Where off-the-shelf AI works, and where it can’t

  • Coding is the obvious “outsource it” case: “I don’t really care if the Python code that Claude Code writes has my unique experience and wisdom behind it.” Same for data analysis and finance workflows — tasks that don’t require your unique context.
  • The opposite list: interviewing a candidate, talking to an investor, talking to a customer, writing to your team. Those depend on taste, judgment and accumulated context that a model trained on the whole internet does not carry.
  • The missing piece he names is the horizontal layer across all of it. Agents today make you more productive and then leave you micromanaging more agents — no layer holds the knowledge of your life together.
  • His framing: the goal is not to replace human interaction but to let you skip to the part that matters. Offload the knowledge transfer, keep the human part.

Why he walked away from a $21M company

  • As CTO of Macro he managed ~15 engineers and became “a glorified information router” — taking information from sales to engineering and back, “a toll booth in the loop.”
  • The deeper reason: both grandparents died a couple of years ago and their knowledge survived only as texts, voicemails and fallible memory in the family’s heads. He wanted to be able to ask his grandfather, an entrepreneur, how he got through the early struggles.
  • That second motive turned out to be a market: 20-40 people he has started working with fall into the legacy category — founders and owners thinking about handing a business on.
  • Concrete user: Pat, a high-end jeweler with a ~20-person store, who now has employees run outreach past his Sentience before sending it, and is building a knowledge base so his daughter can ask how he thought about advertising, or about a tough customer, and get concrete examples back.

Dogfooding: he is the strongest power user

  • Mid-interview his Sentience fielded three questions from his team and answered all three on Slack while he stayed fully present on the call.
  • Engineering standup runs entirely through a standup skill: it iterates each person like a real standup, pulling context from Linear, GitHub conversations and code, then emits the update. Humans review the simulation together and go straight to the two or three sticking points.
  • The tell that it’s working: sometimes two Simulated teammates disagree, or one cracks a joke the way a teammate would. That’s the moment it stops being a generic agent.
  • He’s not sentimental about standups. He has never met an engineer excited about them; they are a necessary evil that exists to solve knowledge transfer — so let the agents do that and spend the 15 minutes on getting to know each other.

Sending the sentience in after the call

  • He interviews 10-15 candidates a week and always runs out of time for questions, so he hands over his Sentience at the end of the call.
  • Roughly two-thirds of candidates then have conversations of 10-50 messages with it; some spend hours.
  • Two things surprise him: he has scaled past his own physical capacity, and candidates say more to the agent than they would to him face to face — because there’s no one to be embarrassed in front of.
  • It has closed contracts and investor commitments mid-conversation. An escalation feature flags the two or three nuggets worth a human follow-up so he can pick the thread up in person.

The master plan: perception, then memory

  • Phase 1 — perception, in three buckets: latent data already in systems you use (Notion, Linear, email, Slack), sensory input from apps that let you deliberately snapshot or bookmark a moment (Apple Watch, iPhone, desktop), and synthesis across modalities into a unified understanding of the entities in your life — you, your podcast, the people you’ve met.
  • No always-on capture. He explicitly rejected the raw-video-feed version; when the host pushes on how creepy an always-listening watch would be, he concedes the trade-off and says they opted for a curated feed instead.
  • Phase 2 — memory. Throwing everything into a RAG vector DB gets you a working version, but it breaks on temporality and updating: a day ago this conversation hadn’t happened, and tomorrow it will have.
  • Their answer is an entity layer “one more layer up” from a knowledge graph, updated in a nightly sleep phase that synthesizes the day into interlinked markdown topics — an auto-generating Wikipedia of your life.
  • Best of both worlds: synthesized, temporally-updated understanding for ordinary questions, plus raw retrieval for the granular stuff (the email address of someone you met five years ago).
  • His complaint about second-brain tools: they demand hours of manual curation a day. He thinks that work is now automatable, and that people shouldn’t have to jerry-rig it.

Personalization: prompt engineering isn’t it

  • Prompt-engineering a frontier model to “talk like this” can get partway there but won’t make the mission a success — and the labs aren’t going to open their weights.
  • What they actually do: fine-tuning and LoRA on open models, creating a model unique to each person once there’s enough context.
  • He still expects a step-function advance to be needed, whether they make it or someone else does.
  • Phase 7 of the plan is literally labeled “unknown” — the executive-function and emotion layer. Phases 1-4 he considers pure engineering; some phases will need net-new research breakthroughs.

The AI-slop diagnosis

  • His hot take: people aren’t really objecting to AI-written text, they’re objecting to everything sounding the same. If 90% of the tweets you saw were written by one human, you’d block that guy too.
  • The uniformity is a property of the three models almost everyone uses — Claude, GPT, Gemini — not of AI as such. They’re different tools that construct output in fundamentally the same way.
  • Which is the argument for hyper-personal models: a billion people with their own Sentience, rather than one voice everywhere.

His advice, and who he’s hiring

  • “No one can compete with you at being you.” You can’t beat a frontier model on code, and the share of engineers who can is going to 1%, then 0.5%. Lean into what makes you weird instead of away from it.
  • Don’t run someone else’s founder playbook. He argues the best founders weren’t emulating anyone — Zuck wasn’t trying to be someone else.
  • Practical: meditate ten minutes a day.
  • Hiring for ML research, full-stack engineering, a growth lead and a product designer. He weights EQ plus IQ, keeps the company flat, is allergic to bureaucracy — and there’s no LeetCode: they just work with you during the interview.
  • On gating: he approves external requests to talk to his Sentience manually, on purpose. The strongest use cases are internal and intimate rather than “put it in your Twitter bio.”

“Who you are uniquely matters now more than ever, actually, because of this increasing uniformity.”

The engineering-side companion is the Sentience Workshop #5 capture — system prompt as one monster, evals before every change, complexity as a mask.