I built the agents first. Then I learned what they were missing.

Eight builds, 500+ conversations and three pivots, from an AI mentor for students to GeniOS. This is how my thinking changed, and what changed it.

GeniOSReadReasonDelegateProofLearn

8AI products built, one on top of the other

500+Conversations: 300+ founders, 200+ managers

3Pivots, each one forced by what people said

30+Enterprises piloting GeniOS today

Here's what I built with AI.

A running log of the agents and tools behind each build.

Let's build with AI

AI as part of my everyday workflow.

I use AI through the whole work of building a company, to move faster and think harder, without handing over judgment.

Think

Before a week goes into the wrong thing.

Market Research

Competitors, reports and customer conversations, turned into a clear view of what to do next.

Decisions

Challenging assumptions and testing the reasoning before a week of work goes into the wrong thing.

Build

A dummy first, then real code, then proof.

Fast Prototyping

A clickable dummy first, so a design gets judged by using it, before any real files change.

Code & Architecture

Real products, with real tests. Not just vibe coding: systems thinking, architecture and scale.

Testing & QA

Scripts decide if it passes. A green run is evidence; a claim is not.

Ship

What actually goes out into the world.

Agents at Work

Email, calendar and orchestration agents, built on LangGraph, Gemini and Supabase.

Films & Voiceover

Long-form explainer films on autonomous companies, made end to end, from research to final render.

Websites

thegenios.com went from plan to live with AI in the loop. So did this site.

Eight builds, and what each one taught me.

I did not start with a theory. I built something, put it in front of people, listened to what broke, and built the next thing on top of it.

Build 1
AI mentor
50 early users
What I thought
Every student could have a mentor, if AI did the mentoring.
What I built
Personalised roadmaps with RAG, Gemini and Supabase.
What it taught mePeople do not need more answers. They need the next step, and someone checking they took it.
Build 2
Resume agent
1,500+ users in 30 days, 3,000+ resumes
What I thought
Take one painful task and remove it completely.
What I built
Fill in the blanks, get an ATS-ready resume with a score.
What it taught meAn agent wins when nobody has to learn anything to use it.
Build 3
Multi-agent orchestrator
Python, FastAPI, LangGraph
What I thought
Agents fail because they act out of order.
What I built
Mr. Elite: intent, tasks, context, policy, execution, aggregation, learning.
What it taught meOrder and policy make agents safe. They do not make them useful.
Build 4
AI employee
Piloted with 5 startups
What I thought
A company will hire an AI employee the way it hires a person.
What I built
Agents for email and calendar, drafting, scheduling and following up.
What it taught meIt did what you said, not what the company needed.
Build 5
Admin AI employee
Piloted with 1 company
What I thought
If I cannot run my own work on this, nobody else will.
What I built
Email, calendar, writing and documentation in one place. I ran it myself first.
What it taught meMost of the job is knowing what not to do, and agents had no way to know that.
Build 6
Chief of staff
Led to the context brain
What I thought
Maybe the missing role sits above the agents.
What I built
An AI chief of staff on LangGraph, following the work across agents.
What it taught meA chief of staff is useless if it does not know the company.
Build 7
Context brain
200+ managers and executives spoken to
What I thought
So agents need memory of the company.
What I built
Persistent memory that every agent could read from.
What it taught meMemory answers what happened. Nobody was answering what should happen next.
Build 8
GeniOS, the autonomous company
500+ conversations, 30+ enterprises, $250k ARR
What I thought
Software should decide what matters, not wait to be asked.
What I built
Reads the company, reasons, delegates to a person or an agent, checks the proof, learns.
What it taught meThis is the part nobody had built.

How I understood memory was not the product.

We were building a context brain. Then the market started telling us the same thing, from four directions.

Memory is becoming a commodity

Vendors are converging on the same features: vector store, hybrid search, graph relationships, temporal decay and entity resolution. On the standard benchmarks they all land in the same place. When Garry Tan built GBrain in twelve days and open-sourced it, the question answered itself. People will build this themselves.

12 daysTo build GBrain. 5,000 GitHub stars on day oneApril 2026

83 to 92%Where cloud memory systems cluster on LongMemEval and LoCoMoPublic benchmarks

$24MRaised by Mem0, with 41,000 GitHub starsOctober 2025

Retrieval is not a brain

A deal at risk shows up as a shift in tone and slower replies. A launch slipping shows up as dependencies nobody closed. None of that is a fact you can retrieve. People infer it from patterns, and search, however good, does not.

$200MARR at Glean, 100+ connectors. Still search, done extremely wellNovember 2025

0 of 8Memory frameworks with enterprise governance: glossary, lineage, entity resolutionAtlan evaluation, 2026

A model on top of memory is not intelligence

It is the obvious next idea, and it fails for three reasons. Without routing it over-retrieves noise and costs climb. Without per-company thresholds, important means the same thing to a 12-person startup and a 4,000-person enterprise. Without precedent, it re-derives answers the company already found.

A wrapper gives plausible answers. Intelligence gives correct ones.

Production is where it all breaks

Demos give an agent clean input and one task. Production gives it messy threads, many teams and competing priorities. An agent may know how to execute a task. It may not know whether that task still matters.

95%Of enterprise generative AI pilots showed no measurable P&L impactMIT Project NANDA, 2025

700Agents Klarna replaced with a chatbot, before rehiring humans over quality2024 to 2025

CAN$812Air Canada was ordered to pay after its chatbot invented a policy2024 BCCRT 149

What compounds, and what does not

ModelsCommoditize
FrameworksCommoditize
MemoryCommoditizes
IntelligenceCompounds

Memory, a company brain and intelligence are three different things

	AI memory	Company brain	Company intelligence
Belongs to	One user and one assistant	The company	The company, for each owner
Comes from	What was said to the assistant	Email, meetings, chat, files, records	The same sources, linked to goals and open work
Answers	What did I say?	What happened?	What should happen next, and who owns it?
Waits to be asked	Yes	Yes	No
Done means	It was stored	It can be found	The proof shows it happened

What I decided to do, and what I decided not to.

What we do

Sit above memory and read from it, instead of competing with it.
Start read-only, and earn the right to do more.
Learn each company's own thresholds and precedents.
Send one situation with one owner and one date.
Close an item only when the proof shows it happened.
Keep people at the centre: they set direction and approve.

What we will not do

Build another memory layer.
Build another agent that writes the same draft.
Send anything without a human approving it.
Trust a confident answer because it sounds right.
Measure success by how good the output looks.
Replace people's judgement with a dashboard.

How I think the next few years play out.

01Every team runs agents
Sales agents draft, support agents answer, finance agents reconcile. Most of the doing moves to software.
02The data stays spread out
Companies keep adding tools, and each agent sees only its own corner.
03A thinking layer sits above them
It knows the goals, reads across every system, and decides what matters today and who acts.
04People manage the work
They set goals, approve actions and correct the system. Each correction makes the next decision better.
05Companies measure autonomy
How much work runs from start to proof with a person only approving, and how often a person has to step in.