Algorithm engineering and server-side systems, with backend development next to them. Each project below ships its tests in its repository, with its decision log or its pre-registration, and every number on my site traces back to a project repository.
- push-lab · page An offline lab for push notification recommendation: recall and ranking checked on public data, send volume and timing tested only in a simulator, with every result labelled by how far it can be trusted without live traffic. On Duolingo's randomized notification logs, the template bandit beats uniform choice by 0.00174 (0.00126 to 0.00221) in two-hour completion per notification (Tier A, randomized real data). In a registered long-horizon study in the simulator, the estimator SMIS, on the project's exploration logs, favours one push a day over none at 6 of 8 settings, the reverse of the truth (Tier C, simulator only).
- ope-checks · page Pre-registered checks of published off-policy evaluation results. On the two confirmatory campaigns of the Open Bandit benchmark, the IPS estimate pooled over both is 0.974 of the on-policy click rate, with an interval of 0.915 to 1.036 over ten-minute blocks: no shortfall shown. A further study re-scores two published estimator comparisons against a reference unbiased for the evaluated target: both replications are reproduced; on All, with ten-minute blocks and the target applied to its own (item, position) pairs, MIPS (w/ SLOPE) against IPS "flips" (closer in 91 of 500 runs); and on Men's, with ten-minute blocks, at the full size of the data and with the reward model and the target fixed, BLUE's unbiasedness is "contradicted".
- retrieval-lab · page BM25 and dense retrieval in Go. BM25, exact dense search and fusion are each accepted against a number that Pyserini publishes for BEIR. BM25 equals the published nDCG@10 and Recall@100 to four decimals on six datasets. BGE-base with exact search is within 0.0006 in nDCG@10 on all six, and within 0.0004 on all but TREC-COVID.
- seqrec-lab · page A pre-registered study of Transformer sequential recommenders, SASRec, BERT4Rec and gSASRec, with equal tuning budgets, five to ten seeds per arm and full-ranking evaluation on one rented consumer GPU. On MovieLens-1M, Video_Games and Steam, SASRec with full cross-entropy is ahead of BERT4Rec and gSASRec, and SASRec with BCE and one negative trails it by more than 2%. MovieLens-1M's test targets had been read for anchor checks before its final runs. MovieLens-20M has no results yet.
- recserve · page A recommendation serving system in Go: retrieval, features, ranking and rerank behind one HTTP API, with the item vectors on gRPC shard replicas. In kind at 100 requests/s, through rolling restarts of the gateways and of the replicas and a bundle swap, no response failed or was degraded, with stage budgets of 100 ms.
- matching-engine · page A Go limit order book with a write-ahead command log and open-loop load testing. The matching core runs 6.25M to 8.88M commands/s on one symbol on a noisy cloud VM, and p99 is 1.76 ms at 10k requests/s and 41.1 ms at 50k, near saturation.
vault-share (end-to-end encrypted file sharing) · backtest-lab (overfitting checks for backtests) · local-ledger (an encrypted personal finance desktop app)