Skill map01
Each row points to the section below that shows it.
| Skill | Where | Proof |
|---|---|---|
| Tracking and registry | SeismicSoCal | MLflow runs carry the commit, data hash and seeds; the @champion alias marks the live model |
| Gated promotion | SeismicSoCal | six statistical and replay rules; a rule that can't run is logged SKIPPED instead of passing |
| Scheduled retraining | SeismicSoCal, BayRidership | a Dagster retrain job; LightGBM retrains every Monday and writes a run log |
| Drift monitoring | SeismicSoCal | a daily Evidently check per station, with an email after two drifting days |
| CI/CD | SeismicSoCal, BayRidership | GitHub Actions on every change; keyless Workload Identity Federation to Cloud Run |
| Infrastructure as code | BayRidership | Terraform for every GCP resource, budget alerts included |
| Load and smoke tests | BayRidership | an image smoke-tested before it's pushed; Locust p95s under load |
| Selection protocol | Berkeley Lab | rolling-origin CV; the test window never picks the model |
SeismicSoCal · QuakeOps: retraining behind a gate02
A live earthquake detection system on 19 seismometers across Southern California. QuakeOps is its MLOps loop: a retrained model can replace the live one only by proving, statistically and on replayed days, that it's at least as good.
- 01
Built the retrain flow: data appended monthly, the challenger trained and logged to MLflow with its commit, data hash and seeds, the champion re-scored on the challenger's unseen test split, then both replayed on 10 held-out days. If the gate passes, the @champion alias moves and the VM pulls the model and verifies its sha256.
- 02
Wrote the gate as six rules: beat the classic baseline, stay non-inferior to the champion, pass the replay, pass the tests, have clean lineage, and have a reason to switch. A rule that can't be evaluated is logged SKIPPED, never a silent pass, and the gate never edits the live thresholds.
- 03
Watched for drift: the daemon logs three scale-free features per station every 30 s, and a daily Evidently job marks each station ok, watch or drifting, with an email after two drifting days in a row.
- 04
Deployed with GitHub Actions (ruff, pytest, the daemon selftest, the site build) and a tar deploy to the VM; systemd respawns the stream daemon if it drops. Live pushes stay off in shadow mode until confirmed precision on live data clearly beats chance.
- 05
Where it stands: the registry, daily pull, drift logging and CI/CD deploys are live, and the gate passed a dry run (champion vs. itself). The first monthly retrain hasn't run yet.
BayRidership · shipping and retraining on GCP03
A Bay Area transit analytics platform I built end to end. Its API and agent run on Cloud Run, its station forecasts retrain every week, and everything in GCP is code.
- 01
Set up CI on every push to main (ruff, pytest, Spark, browser, container and Terraform jobs) and keyless deploys: Workload Identity Federation pinned to the repository's main branch, with no service-account keys in GitHub.
- 02
Wrote every GCP resource in Terraform: BigQuery datasets, the bucket, Artifact Registry, Cloud Run, least-privilege service accounts, and $1 and $5 budget alerts. The agent's Gemini key lives in Secret Manager.
- 03
Ran the API on Cloud Run with 0–2 instances, scaling to zero, from a multi-stage non-root image that is smoke-tested before it's pushed.
- 04
Retrained the LightGBM forecasts every Monday; each run logs MAE, RMSE and coverage with their CIs, and the API picks the band's label from that run's own coverage.
- 05
Backed it with tests: 179 pytest, 52 node and 75 Playwright tests, 84 dbt data tests, and reproducibility checks that recompute the A/B, the causal study and the trend adjustment from committed inputs.
Berkeley Lab · a model-selection protocol04
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
Trained on a chronological hold-out with early stopping on validation loss.
- 02
Selected the lookback with rolling-origin cross-validation: train on the past, forecast the next block, slide forward. Persistence was scored on the identical windows, and the test window never picks the model.
- 03
Re-ran the chosen model on 5 random seeds. Its error moved only slightly, so the choice is a stable finding rather than a lucky initialisation.
- 04
Kept the pipelines reproducible: every site is loaded, cleaned, resampled and plotted the same way, and the whole thing reruns end to end on new data.