Skill map01
Each row points to the section below that shows it.
| Skill | Where | Proof |
|---|---|---|
| Framing a testable question | CBU Research | is the R² ≈ 0.10 ceiling a model problem or a data problem? |
| Baselines from the field | SeismicSoCal, Berkeley Lab | STA/LTA and amplitude + distance for quakes; persistence for traffic |
| Ablations | SeismicSoCal | nearest-station ablation: magnitude R² 0.951 → 0.808 |
| Hypothesis tests on failure modes | Berkeley Lab | five losses and percentile forecasts, both rejected |
| Variance vs. luck | SeismicSoCal, Berkeley Lab | a 10-seed study; seed re-runs of the chosen lookback |
| Pre-registration | BayRidership | design hashed before any result; a 200-simulation check |
| Time-series Transformers | Berkeley Lab | PatchTST: patches as tokens, RevIN, channel independence |
| Mentoring research | AperioHub | one-to-one research that ends in a defensible result |
CBU Research · a data limit, shown with a paired design02
ML Research Intern at California Baptist University, May 2026 – present.
An inverse-FEA model (predicting a structure's stiffness from its measured shape) was stuck at R² ≈ 0.10 on one target, holding up the team's IEEE paper. I asked whether that was a model problem or a data problem.
- 01
Made the evaluation trustworthy first: leak-free scaling, out-of-fold scores, a predict-the-mean baseline and a measured overfitting gap, so a small difference couldn't be an artefact.
- 02
Benchmarked seven configurations with different inductive biases under the same cross-validation. If the model were the bottleneck they would disagree; they all hit the same ceiling.
- 03
Ran a paired clean-vs-noisy design, then ruled out the obvious bottlenecks (PCA, the noisiest inputs) with a bypass and an ablation.
- 04
Tested the realistic deployment, trained on clean simulations and used on noisy measurements, and swept injected noise to draw an identifiability curve: how precise the measurements must be for each target to work.
- 05
Wrote it up for the team's IEEE paper. The figures and detailed results are held back until publication, so only the method is public.
Berkeley Lab · PatchTST and the limits of history03
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. I test how far history alone can predict that traffic.
- 01
Implemented PatchTST, a Transformer that cuts a series into short patches and treats each one as a token. Attention cost grows with the square of the token count, so patching is what makes a long lookback affordable. RevIN normalises each window, and the head predicts every future day at once, so errors don't compound.
- 02
Kept channel independence and stated its trade-off: each traffic variable goes through the same network separately, which multiplies training examples and curbs overfitting, but the model can't know that hits plus misses must equal accesses.
- 03
Swept the lookback per horizon with rolling-origin CV: about three months of history did best a day ahead, while at 30 days short windows won. A grid on the finalists and re-runs over seeds showed it wasn't luck.
- 04
Tested whether the missed peaks were a training problem (five losses, then forecasts aimed high) and rejected it. That changed the team's question from which model predicts peaks to what signal would make them predictable.
- 05
Split calm and erratic eras of one cache: the best amount of history depends on the regime.
SeismicSoCal · deep vs. classical, then a replay04
A live earthquake detection system on 19 seismometers across Southern California. I tested whether this live system would have alerted correctly on real days it never saw.
- 01
Designed two architectures: a CNN → Transformer that reads one station's 30 s window to detect, and a CNN → GNN → Transformer over the station graph to size, where the CNN reads waveform shape and the graph reads size.
- 02
Bootstrapped confidence intervals by event; the deep and classical intervals don't overlap on either task.
- 03
Checked the detector doesn't depend on where the quake starts: AUC 0.9997 / 0.9999 / 0.9997 with the P-wave at 2, 12 or 22 s.
- 04
Ran a 10-seed study: ensembling added +0.003, inside the sampling CI, so the live system keeps 5 seeds.
- 05
Made a replay the acceptance test, with thresholds frozen on validation days and test days scored once. Swarm tuning followed a rule fixed in advance and raised M3+ catch from 71% to 78%.
BayRidership · pre-registered studies05
A Bay Area transit analytics platform I built end to end. It includes a causal study of the eBART extension and an A/B test of its SQL agent, both written down before any result was seen.
- 01
Committed the design (station sets, windows, bins, the seed, the verdict rule) as a hashed file before computing any treated-vs-control number.
- 02
Used placebo inference: give the extension to every set of 3 control stations instead and re-run. The real line is more extreme than all but 3 of them.
- 03
Validated the method on 200 simulations first. That caught a too-narrow interval, which one dated amendment fixed before any real estimate.
- 04
Stress-tested the finding: exits instead of entries, 16 suburban-only controls and a monthly 2016–2020 series agree in sign, and a fake opening date finds nothing. The post-hoc trend check (+1.0%) is reported and labelled.
- 05
Ran the agent A/B the same way, with a power analysis before arm B; a
--checkcommand recomputes every number in the write-up from committed summaries.
AperioHub · mentoring research06
Part-time Research Mentor in AI & ML at AperioHub Advisory, Oct 2026 – present: a one-to-one, project-based research engagement with a high school student.
- 01
Every session runs the same loop: review what was done and what it showed, introduce one new concept, assign one focused task.
- 02
The direction is chosen with the student from five: model evaluation, neural networks, open datasets, LLMs, and AI ethics and bias.
- 03
I teach the habits from my own research: baselines and held-out tests, pre-registration, knowing when the answer is a ceiling, and plain-language method write-ups.
- 04
I record each session, and AI-assisted notes keep a running record of the project, dead ends included.