Skip to index

Hussam AhmedStaff Technical Program Manager Building large-scale systems by day, exploring the universe by night.

By day: autonomous driving and the AI platforms underneath it — the architecture between teams, and the interfaces they have to agree on. By night: a telescope, and the software that decides where to point it.

Six horizons below · 100 ms → 23 Myr · day first, night last

The day half starts small. A vehicle gets about a hundred milliseconds to decide, working from whatever its sensors happened to catch.

Then I have to be able to explain it.

A safety case wants the reasoning, not just the output. So does an incident review six months later, and so does the engineer deciding whether to trust it. Most of that work sits in how the system is put together rather than in the model.

Sensors see geometry. The Map Knowledge Graph adds judgment.

That was one scene. I work with a few hundred thousand.

Six cities worth, and the question stops being whether a model works. It becomes how you know it still works after the next change, on roads nobody has driven yet. Coverage sets, evaluation harnesses, regression suites, release gates — this is where most of my time actually goes.

By the time it reaches me, the software is the tractable half.

Most program risk is not technical.

It is two teams holding different assumptions about the same interface, and nobody noticing for a quarter. Dozens of groups, each doing something locally sensible, adding up to something that does not integrate.

My job is catching that early, putting the tradeoff in front of whoever owns it, and keeping one plan the whole program can work from.

And the teams are rarely in one place.

A program on three continents runs on written interfaces and explicit ownership, because there is no room where everyone is present. Anything that stays unwritten gets decided separately in each place, and you find out two months later.

And this is the other half.

Thirteen hundred objects are up tonight.

Altitude, moon separation, how long each one stays clear of the roof, whether it fits the sensor at all. The constraints cut it down to one. I wrote the software that does the cutting, mostly so I would stop doing it by hand at midnight.

M51 — the light left it twenty-three million years ago and lands as about four and a half hours of exposure. It is the only system I build where I am also the entire team.

Staff Technical Program Manager

Hussam Ahmed

I work on systems outside controlled environments — incomplete data, moving requirements, and constant production pressure.

That means dealing with trade-offs and conflicting priorities while still delivering something reliable.

This site breaks down how I design, execute, and solve problems in that context.

What I work on

Technical program leadership

  • Large-scale, multi-team engineering programs
  • Architecture and interface governance
  • Risk, dependency and critical-path management
  • Distributed teams across time zones

AI systems & platforms

  • Production AI systems in reliability-critical settings
  • Model evaluation, guardrails and release gates
  • Agent platforms and AI-assisted engineering workflows
  • Turning capable models into shippable systems

Autonomous driving

  • Perception, prediction and planning interfaces
  • HD maps and queryable driving context
  • ODD definition and feature-activation gating
  • Explainable decisions for the safety case

Selected work

  1. Map Knowledge GraphAn HD map turned into something a vehicle can query — lanes, signals, occlusions and priors linked into decisions it can defend. Seven scenarios, every state addressable.
  2. AI Knowledge HubEighty-nine interactive guides to working with coding agents and AI platforms, verified against primary sources rather than assembled from posts about them.
  3. Knowledge Defined VehiclesWhy map geometry alone is not enough, and what changes when driving context becomes queryable. The argument behind the demonstrator.
  4. BlogBuild logs from systems that had to work — what broke, what the real constraint turned out to be, and what survived contact with production.
  5. AstroBrainThe observatory planner: one real night computed from the ephemeris, 1,309 candidates culled to one, and the exposure arithmetic behind the decision.

Latest writing