I build systems where correctness is checkable: payments with compensation flows, LLM pipelines that treat model output as untrusted input, and robot-learning experiments with honest measurement. Pick your section below; several projects appear in more than one, because real systems rarely fit one label.
Streaming two-pass ETL over 1.31M Government of Canada contract records; JWT auth and RBAC; an append-only audit log enforced by MySQL grants and proven by integration tests; six parameterized SQL red-flag indicators, including a CTE and a date-windowed self-join.
Two independently deployable services with per-service data ownership, BigDecimal money, optimistic locking, and a compensating refund flow when a credit fails after a debit.
The analytics platform behind my Summer 2025 data engineering internship: ETL ingest of 500+ estimates (~$7.4M tracked revenue) into Azure SQL, typed T-SQL view layers, and a Blazor reporting dashboard, provisioned as infrastructure-as-code.
Every numeric token in a generated brief must resolve through a JSON path into the evidence pack; every red flag must cite a fired SQL indicator; a failing draft is re-prompted with the exact reasons, up to 3 attempts, then falls back to a deterministic brief with provenance tags. An adversarial red-team suite gates CI on every push.
The LLM may only propose JSON against a schema; a hand-rolled validator checks every field against the live command vocabulary; exactly one corrective re-prompt quotes the precise error back; a second failure falls through to a deterministic grammar parser, with per-command provenance in the UI.
A claim-by-claim reproduction of the ACT paper (RSS 2023) with Wilson confidence intervals, disclosed confounds, explicit not-tested scoping, and an errata section correcting my own earlier numbers.
Imitation learning (ACT, 90% from pixels), VLA fine-tuning (SmolVLA 0→90% language-conditioned on a free T4), classical 6-DoF grasping from first principles (20/20 picks, 0.58 mm), RL with a from-scratch world model and CEM-MPC planner, and a ROS 2 deployment with SLAM + Nav2 and a C++ port at 0.000 mm parity.
Fleet supervision for simulated AMRs: hand-rolled rosbridge v2 client with reconnect and subscription replay, a 60 fps interpolated canvas map over a 10 Hz backend, and per-robot emergency-stop lockouts.
On-device gesture recognition in the browser with a crane-signal vocabulary adapted from ASME B30.5, driving a simulated cobot through a safety state machine with dwell, debounce, and refractory stages.
Hugging Face LeRobot — root-caused and fixed an unsafe video-backend selection bug that crashed dataset decoding on Windows. My fix was merged upstream as a co-authored commit after maintainer review (PR #4307, which supersedes my #4177). A second fix for Hub path normalization is under review (PR #4174), plus independent Windows verification of #4121.