AI systems · Lenovo → now
Building the engineering around the model
The model is the easy part. The system around it is the work.
Context
Applied AI work at enterprise scale, and then again from scratch inside my own companies. Two very different budgets, the same architecture questions.
The problem
Most AI projects fail somewhere other than the model: data movement, evaluation, cost per request, latency budgets, and the operational reality of running inference continuously rather than in a demo.
What I built
AI systems architecture and hands-on prototyping — the pipelines, the serving decisions, and the evaluation loops that decide whether a model is actually usable in production. I still write this code rather than review it.
Why it mattered
It moved AI from something demonstrated to something operated, with the same discipline the rest of the infrastructure gets.
Technical decisions
- Treat inference as a capacity problem, not a feature — it has a throughput curve like any other workload.
- Design the evaluation loop before the model choice; the loop outlives every model you swap in.
- Keep the data path short. Most latency and most risk live in the movement, not the math.
Why it matters now
Every company I'm building runs on this pattern. The frontier moves; the engineering questions underneath it barely change.
- AI
- Systems architecture
- Inference
- Evaluation