AI/ML Systems and Software Engineering Projects

Problem

Recruiting updates arrive through email, while trackers and follow-up plans drift without bounded evidence and review.

What I Built

Built Gmail triage, typed workflow extraction, confidence-aware review, deterministic fallback, and model traces over 130 versioned sanitized fixtures.

Impact

The deterministic product-loop regression passes 118 of 130 contracts (90.8%) with 0 unsafe automatic mutations; 12 false review routes remain documented in checked-in error analysis.

Problem

Serving comparisons become misleading when request modes, artifacts, hardware evidence, or run completeness differ.

What I Built

Built streaming and non-streaming clients, concurrent open-loop and closed-loop workflows, JSONL/CSV artifacts, manifests, latency reports, KV-cache estimates, quantization comparisons, optional GPU telemetry, and claim-audit gates.

Impact

The toolkit rejects mock, incomplete, or incomparable runs from performance rankings and preserves the evidence needed to audit each comparison.

Problem

Tool-using agent experiments need reproducible tasks, isolated execution, and controls that expose harness failures.

What I Built

Built 25 terminal, SWE-style, and stateful-tool tasks with deterministic grading, replay, read-only container roots, dropped Linux capabilities, seccomp, no-new-privileges, PID limits, and explicit writable mounts.

Impact

The scripted oracle and deterministic local control each score 25/25; the no-op negative control scores 0/25.

Problem

Agents need fast local recall without replacing canonical files or leaking secrets across workspaces.

What I Built

Built SQLite FTS5/BM25 retrieval with provenance, source snippets, freshness verification, secret redaction, workspace filtering, and CLI, Python, and stdio interfaces.

Impact

On 120 sanitized synthetic queries, the benchmark records 0.990 Recall@1, 0.995 MRR, 1.000 no-answer precision, and 0 privacy leaks; exact-phrase grep reaches 0.200 Recall@1.

Problem

Sequential recommendation models need a shared split and metrics before their ranking quality can be compared.

What I Built

Contributed to preprocessing, model implementation, training, evaluation, comparison, and reporting across 23,936 users, 2.62M interactions, and 50K items with leave-one-out evaluation at K=50.

Impact

SASRec reaches HR@50 0.1525 and NDCG@50 0.0503, about 4.8x and 5.4x the MostPopular baselines.

Problem

Xiangqi research needs a correct rules engine and reproducible training and evaluation workflows before model strength can be studied.

What I Built

Built an 8,100-action move space, legal policy masking, a PyTorch policy-value CNN, PUCT MCTS, self-play, checkpoint training, engine distillation, and CPU, MPS, and CUDA evaluation workflows.

Impact

The repository provides a reproducible baseline for rules, search, training, and engine comparison work.