I design and ship agentic systems for teams that need them to be reliable, not just impressive. The approach comes from doctoral research in neural networks, years of forecasting under uncertainty, and leading teams that put models into production.
Most agent projects stall between the demo and the second week of real use. The work that closes that gap is mostly unglamorous: scope, grounding, evaluation and clear hand-offs to people.
Planning, tool use, memory and hand-offs designed around one narrow job at a time. I start from the workflow and the cost of a wrong step, then decide how much autonomy the system has earned.
Agents that act on your documents and data, not on what a model happens to remember. Every claim and every action can be traced back to the source that justified it.
Test suites for behaviour, not just accuracy: task success, tool-call correctness, cost per run, and the failure cases that matter to the business. Measured before launch and watched after.
Clear points where the agent stops and asks. Approvals, escalation paths and audit logs that let a team trust the system without having to watch it all day.
None of them are new. They come from forecasting and research, where a confident wrong answer is the most expensive kind.
Time series, then language, then systems that act. The tools changed. The question stayed the same: what is the data actually saying, and how much of it can we honestly stand behind?
Doctoral work on neural networks for forecasting, most recently spiking neural networks: models that fire the way biological neurons do, sparingly and only when something is worth signalling. I have presented that work outside my own field, including to physicists at a statistical physics conference, on what a machine can learn about the shape of short-term memory.
Long before generative AI I was modelling what happens next: demand curves, churn, anomalies buried in years of sensor data. The habit stuck. Systems should state their uncertainty out loud instead of hiding it behind a single number, and agents are no exception.
Retrieval-augmented pipelines and language models wired into workflows where a wrong answer has a real cost. Production systems, not demos. This is the ground the agent work stands on.
I have led engineering and data science teams through that whole path, translating between what a model can do and what a stakeholder is willing to trust. Roadmaps, not just repositories.
The best collaborations usually start with a question, not a pitch.