An explorable profile

Most pages tell you what someone did.
Poke at how I think.

Think. Model. Build. Ship.

I'm Pouyan Jahangiri — AI architect and engineering leader, with roots in mechanical engineering and applied math. I build AI that survives contact with production: retrieval, knowledge graphs, agentic services, and the ML platforms under them — recovering the hidden structure in messy real-world signal and shipping it.

consulting & quick ML / AI advice
Portrait of Pouyan
01 / the core instinct

Find the nearest meaningful thing in a space of millions.

Retrieval, recommendation, RAG — under the hood it’s one question: given a point, which others are closest? Here are 120 points. Grab the orange one and move it.

5
The orange node is your query. Lines connect it to its 5 nearest neighbors, recomputed live as you drag. Toggle the partition to see the kd-tree-style splits that let this run in log time instead of scanning everything — the same recursive bisection behind ANNOY-type indexes I've built teaching tools for.
02 / a patented idea, made tangible

Forecasts that agree with themselves.

Predict a company’s total, and predict each region — they rarely add up. My patent reconciles the levels. Drag the trust dial: at 0% you trust the regions; at 100% you force them to match the top-line total.

0%
Thin lines: per-region forecasts. Teal: the official total. The dashed line is the sum of the regions — watch the gap close as you reconcile. Hierarchical, coherent, and provably no worse than either level alone.
03 / many small minds, orchestrated

Break a hard task into agents that find their own way there.

A deep-agent harness doesn’t answer in one shot — it decomposes. Each step can spawn 3 sub-agents and recurse 2 levels deep. Drag across the tree to follow one trajectory, from the task down to a tool call.

3
2
Grey nodes are sub-agents the harness spawns; teal leaves are tool calls. The terracotta path is one trajectory from task to tool — drag across the figure to follow a different one. More branching and depth means more coverage and more cost; a good harness decides how far to go.
04 / intelligence at the edge

Sometimes the smartest move is to never call the cloud.

On-device inference skips the round trip — but a small device is slower per FLOP. Set the network round-trip to 60 ms and drag the model-size marker. Left of the crossover, ship it to the edge.

60 ms
Terracotta: total latency on-device, which climbs with model size. Teal: cloud — a fixed network round-trip plus faster compute. Left of the crossover the edge wins; drag the marker to read both off, and scrub the round-trip to move the line. The basis of the sensor-to-SaaS systems I build.
05 / where this came from

The same move, at every scale of my career.

Finding hidden structure started literally — modelling particles at a national accelerator — and kept reappearing: in knowledge graphs, in forecasting, in agentic systems, and now in building companies around them.

VP of Data & AI — Fenelagentic CRM · 2026 →
Founder — MedBrain Technologieshealthcare AI · PatientCraft
Director / EM — ODAIA Intelligencepharma AI platform
EM / Lead — Cymax Group Technologiese-commerce platform
Research Engineer — TRIUMFnuclear medicine & engineering PhD & MASc, Applied Science & Engineering — UBCphysics