Senior research engineer, focused on the platform.
I build the framework, tooling, and platform that turn a quant's hypothesis into a monitored production strategy. I like the parts of the loop that get less attention — the registry, the feature graph, the deploy contract, the observability — because that's where reproducible alpha actually lives.
The site is a working artifact rather than a description of one: Research holds the program census and the writeups, the sandbox runs an ML pipeline in the browser on synthetic data, the backtest hits the live alphakernel HTTP service, and /agents shows the AI-research-agent posture (typed operator, MCP tool surface, citation refusal).
It also runs as the operator workspace — the Now hub is the morning briefing, /strategies is the catalog, /features is the leaderboard — with the platform underneath documented in System (ADR-0059).
How I work
-
The framework is the leverage.
A research team's velocity is bounded by how cheaply they can ask the next question. Most of my work is removing the cost from "next question".
-
Point-in-time is a property of the framework, not the user.
If a researcher can leak the future into a model by accident, the framework lost. The boundary is enforced where it can't be skipped.
-
Two ownerships, two pagers.
Freshness pages on call. Alpha decay surfaces to research. Mixing them costs you sleep without buying you safety.
-
Reproducible means bit-exact.
Every artifact carries enough lineage to be re-run today, next quarter, or three years from now under audit.
-
Promotion needs two answers.
A number that beats the incumbent, and a sentence that explains why. The framework asks for both.
Background, in one paragraph
A decade across full-stack engineering, data, and ML infrastructure — most useful when those three meet. STEM background. Equally happy in Python and the runtime layer underneath it.
A working method
When I introduce a new piece of platform, I run a two-week pairing program with the two heaviest users first. They don't read docs; they push code. I watch which parts they fight with, and those become the API. Adoption isn't a roadmap line item — it's the natural consequence of the framework-only path being faster than the alternative.
If you want to talk
Questions about a technique, a kill, or how a surface is wired are welcome — the fastest route is [email protected]. If a result here looks wrong, say which page and which number; every claim on the site is meant to be checkable, and one that isn't is a bug.