Skill map01
Each row points to the section below that shows it.
| Skill | Where | Proof |
|---|---|---|
| Deep learning | SeismicSoCal | a CNN → Transformer detector and a CNN → GNN → Transformer sizer |
| Training for live conditions | SeismicSoCal | live-noise hard negatives; epicentre, pick and station-drop jitter |
| Forecasting at scale | BayRidership | one global LightGBM per quantile, 50 stations, 1 to 14 days ahead |
| Calibration | BayRidership | a split-conformal band whose label changes only when a back-test earns it |
| Time-series Transformers | Berkeley Lab | PatchTST in NeuralForecast, lookback chosen by rolling CV |
| ML in a product | Academy of Testers | a BKT + IRT engine as pure functions behind Spring Boot |
| Testing ML code | Academy of Testers, SeismicSoCal | 36 JUnit tests; pytest and a daemon selftest in CI |
| Serving and promotion | SeismicSoCal | every 2 s per station on a free ARM VM; a six-rule promotion gate |
SeismicSoCal · two models serving a live stream02
A live earthquake detection system on 19 seismometers across Southern California. Two PyTorch models run on a free ARM VM, scoring every station every 2 s.
- 01
Built the detector as a CNN → Transformer over a 30 s window, with the P-wave anywhere in it and 2,062 hard negatives: real live-stream noise that once fooled a detector.
- 02
Built the sizer to separate shape from size: each window is normalised to unit peak so the CNN reads shape, while log peak velocity and distance are graph-node features. Two graph-convolution layers (Gaussian distance weights, σ = 50 km) and a Transformer across stations feed the head.
- 03
Trained under live conditions: the epicentre jittered by ~8 km, picks by ±0.5 s, and stations dropped at random.
- 04
Kept 5 seeds rather than 10 after a seed study showed ensembling added only +0.003. The same
pipeline.pyruns live and in replay, and checkpoints carry their normalisers and station list. - 05
Put promotion behind QuakeOps: a retrained model goes live only through MLflow tracking and a six-rule statistical + replay gate.
BayRidership · global quantile forecasts, retrained weekly03
A Bay Area transit analytics platform I built end to end. Its forecast covers every BART station two weeks out, with a band that says how often it holds.
- 01
Framed each training row as a station, a forecast origin and a horizon, with features that use only data up to the origin. Values are scaled by each station's 28-day level, so one model fits Embarcadero and the quietest suburban stop alike.
- 02
Trained one global LightGBM model per quantile (p10, p50, p90) instead of one model per station.
- 03
Back-tested walk-forward over 12 origins: the first 6 for calibration, the last 6 for testing. The improvement's 95% bootstrap interval is 200 to 340 entries a day.
- 04
Calibrated the band with rolling split-conformal prediction. It held 73% against an 80% target (the Christmas fold alone is 44%), so the site calls it a “calibrated model range”, not an 80% range.
- 05
Retrained every Monday with a run log of every metric and CI; FastAPI on Cloud Run serves the forecasts.
Berkeley Lab · PatchTST in practice04
Data Science & ML Research Intern at Lawrence Berkeley National Laboratory (ESnet), Sep 2026 – present, forecasting traffic at the regional data caches that serve CMS physicists.
- 01
Ran NeuralForecast's PatchTST on PyTorch Lightning with RevIN on and early stopping on validation loss.
- 02
Made the patch follow the lookback: 16-day patches with an 8-day stride once the window is long enough, and half the window for very short ones, since a patch can't be longer than its input.
- 03
Standardised each variable, because counts and data volumes live on very different scales.
- 04
Chose the lookback by rolling-origin CV, and tested the loss and percentile forecasts as experiments instead of leaving defaults.
Academy of Testers · an adaptive engine in production05
A full-stack AP/SAT study platform I built as ML & AI Software Developer (Jan 2023 – Jun 2026). The adaptive SAT engine and the RAG essay grader both run in the live product.
- 01
Wrote the engine as pure functions with no Spring, no repositories and no
now(): clocks and data are passed in, so all of it is testable without a database. - 02
Damped BKT updates (25% of a gain, 50% of a loss) with a guard so a miss never raises mastery; a miss also bleeds into prerequisite skills at κ = 0.35.
- 03
Applied forgetting lazily on read in
MasteryService, the only place weights are read, so a stale value can't leak out. - 04
Scored candidate questions as 0.45 × skill gap + 0.40 × Fisher information − 0.15 × recency, then sampled from the top 5 with a softmax (τ = 0.15). The client does no model math.
- 05
Ran the essay grader on GPT-4o at temperature 0 with constrained JSON output, grounded in retrieved rubric clauses and scored exemplars.