Evidence,
honestly weighed.
Mathematics → game development → AI engineering.
I like understanding the structure underneath an answer: what it assumes, and whether the evidence actually supports it. A bias for honest evaluation over shiny demos, most recently SteamLens, a live LLM product that cites its own measured accuracy. Built in public.
Featured work
SteamLens
What do Steam reviews actually praise and criticize and how much trust should you place in the answer? SteamLens is an evaluation-first review analysis system planned as four milestones; the first three are built, measured and live, and the fourth is deliberately deferred. It turns unstructured reviews into measured, aspect-level evidence, with the labeler's quality, cost, and failure modes published beside the numbers.
- The Instrument Around the Modelextraction + evaluationThe measuring device before the product: a census labeled for $3.80, and its error measured against human judgment rather than asserted.
- Sampling Without Random Accessthe sampling studyHow few reviews a report actually needs: a size rule validated on games it never trained on, with the distance from uniform sampling measured and priced instead of pretended away.
- The Instrument Goes LivedeploymentThe measured pipeline behind a public URL: narrated live analysis, a fenced narrative, spend gates, and a deploy pipeline that ships only what a human approved.
The Layer Under the Products
Two deployed projects, two hosts, one edge, and one shared operational layer between them. It appeared when the second project's ingress changes started landing in the first project's repository. That layer was extracted into its own repository behind an explicit ownership line, then held to one test per claim: what existed adopted on zero-diff plans, the host rebuilt from a blank machine by a playbook, and a rebuild-and-restore drill from nothing in about an hour.
Blackjack RL
A four-part research arc: a measurement instrument that had to earn trust first, then one question asked three times with the ground truth progressively removed — can learning rediscover optimal play as a table, as a network, and finally as a bet-sizing rule whose signal is buried fifty-deep in per-hand noise?
- Monte Carlo Blackjack Simulatorthe instrumentA self-validating engine, checked against published blackjack mathematics — ~90M hands that every later claim is measured against.
- The Policy Audittabular controlTabular Monte-Carlo vs. the proven-optimal table: ~93% rediscovered, and the missing part traced to coverage, not method.
- From Table to NetworkDQN vs. tableA network repairs the table's coverage gaps — and pays for the smoothness exactly at the policy's sharp boundaries.
- Betting Against the NoiseKelly vs. RLThe capstone: the edge is real and measured — and provably too thin, rare, and noise-buried for end-to-end learning to hear.
Learning to Guess the Distance
A learned A* heuristic that looked like a wash — until the pooled average was split and two opposite effects appeared. One regime-tag feature turns it into 17% less search at a 0.2% cost; moved off its training distribution, it fails in two opposite ways. Two axes, never collapsed into one score.