Research
Dependable intelligence in real systems
One research program, not several. Everything below is an attempt to answer two questions about what it takes for probabilistic intelligence to survive contact with the real world. I publish the questions first, then the experiments, then whatever the experiments actually say — including the negative and unfinished ones.
Question 1
Dependability
How do probabilistic AI systems become trustworthy, observable, reproducible, and reliable?
determinism and consistency · replicability · verification and evaluation · observability · calibration · knowing when not to answer
The thread that connects the agentic-platform architecture I designed at Apple to what enterprise builders kept asking about in Bangalore. Agents don't fail because the model is weak; they fail because the system around it was never legible enough to verify.
Question 2
Intelligence under constraint
What changes when intelligence has to run under latency, privacy, cost, energy, offline, and physical constraints?
small and local language models · edge inference · privacy and data locality · latency, cost, energy · offline operation · cloud/edge routing
A model running on the phone in a student's hand, without connectivity, is a different research object than a frontier model behind an API. The interesting part is the seams — what stays local, what escalates, and how a system knows the difference.
Testbeds
Where the two questions get tested. These are applications, not separate research programs — a testbed earns a question of its own only when finished work justifies it.
- Small models on real devices — The Acropolis Edge Lab's two phone-only projects — offline Hinglish voice Q&A and on-device campus computer vision — are the first place both questions get measured at once.
- Physical and geospatial systems — A decade of maps taught me how hard the physical world is to represent, and how unforgiving it is about being wrong. Cameras, sensors, and imagery reopen the question at a smaller scale.
- Agentic systems — Agents are the sharpest available stress test for dependability: they expose the debt a system already carries — fragmented data, knowledge in people's heads, workflows that exist nowhere in machine-readable form.
Latest work
Findings replace plans here as they arrive. Everything below is published; nothing claims a result it doesn't have.
- 2026-08-18Small AI, Real Systemsworking notes
What a week in Bangalore changed in my thinking — why the questions enterprise builders ask about small AI are the same ones that shaped an agentic platform at Apple.