Skill map01
Each row points to the section below that shows it.
| Skill | Where | Proof |
|---|---|---|
| SQL and data modelling | BayRidership | 17 dbt models on BigQuery: a star schema, an SCD2 station dimension and KPI marts |
| Cleaning and KPI definitions | BayRidership | every dropped row gets one reason and an audit; each KPI is computed once |
| Exploratory analysis | Berkeley Lab | Pandas EDA that found right-skewed peaks, change-points, and calm vs. erratic eras |
| Forecasting against baselines | Berkeley Lab, BayRidership | persistence beaten on every variable; station MAE 48% below seasonal naive |
| Experiments and A/B tests | BayRidership, Academy of Testers | a pre-registered A/B test with McNemar's test; an essay grader A/B-tested against official scores |
| Rare-event error analysis | SeismicSoCal | precision next to a chance baseline; every missed M3+ quake sorted by cause |
| Diagnosing a data limit | CBU Research | a paired clean-vs-noisy comparison that traced the ceiling to the data |
| Building a gold-label benchmark | BearLM | questions written from real chunks, each with an exact answer key |
| Measuring a fix | NumIsToken, Kigumi Group | load time ~2–3 s → ~300 ms; coverage 90% → 98% |
| Teaching the method | AperioHub | baselines, held-out data and leakage checks, taught one to one |
BayRidership · warehouse, KPIs and decisions02
A Bay Area transit analytics platform I built end to end. Eight years of BART and Bay Wheels ridership sit in one BigQuery warehouse, and the questions it answers are an analyst's: how far has ridership come back, did a new extension add riders, and which stations can BART least afford to close?
- 01
Cleaned the raw trip files with PySpark so every dropped row carries exactly one reason (bad date, bad station code, duplicate and so on), with an audit of rows in, rows out and drops per reason. DuckDB merged Bay Wheels' two source schemas into one.
- 02
Defined each KPI once in a dbt mart (recovery vs 2019, rolling 28-day entries, year-over-year change, peak-hour share, busiest station), so the site and the SQL agent read the same columns instead of re-deriving them.
- 03
Found that in 2025 weekday BART ridership was at 43% of 2019, while Bay Wheels ran at 127–210% of its 2019 monthly level, so bikes per 1,000 BART entries rose from 21.1 to 80.6.
- 04
Scored the 15 stations BART may close on four rules shown as text (strong 6, mixed 7, weak 2) and checked the press against the warehouse: “about 12%” of ridership is 12.6%, and South San Francisco's “1,538 weekday riders” is 1,409 exits. Close calls at a rounding edge are reported, not moved.
- 05
Measured the eBART extension with difference-in-differences: compare how the line changed after opening with how 34 far-away stations changed over the same weeks. The line gained +4.2%; a post-hoc trend check says +1.0%, and I report both.
Berkeley Lab · exploring and forecasting cache traffic03
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 explored the data before modelling it.
- 01
Built reproducible Pandas pipelines that load, clean, resample and plot every site the same way and rerun end to end on new data.
- 02
Let the exploration shape the modelling: accesses always equal hits plus misses, the sites behave differently, peaks are right-skewed (most days ordinary, a few extreme), and change-points shift the level over time.
- 03
Looked past the headline RMSE. The first model beat both baselines on most variables but smoothed straight through every spike, so from then on I judged each experiment on both average error and peaks.
- 04
Chose the lookback with rolling-origin cross-validation (train on the past, forecast the next block, slide forward) and scored persistence on the identical windows, so every comparison is like for like.
- 05
Tested two explanations for the missed peaks and rejected both: the signal isn't in past traffic. I present each week's result to the group.
CBU Research · finding a data ceiling04
ML Research Intern at California Baptist University, May 2026 – present.
A model predicting a structure's stiffness from its measured shape (an inverse-FEA problem) was stuck at R² ≈ 0.10 on one target, holding up the team's IEEE paper. I set out to find out whether tuning could help at all.
- 01
Made the evaluation trustworthy before comparing anything: scalers fit on the training fold only, every score out-of-fold, each target checked against a predict-the-mean baseline, and the overfitting gap reported.
- 02
Benchmarked seven model configurations under the same cross-validation. All of them landed on the same ceiling for the hard target, while the easier targets were predicted well by every one.
- 03
Trained the same models on clean and on noisy data, side by side. The target was recoverable from clean data and lost once realistic noise was added, so no amount of tuning could bring it back.
- 04
Swept the noise level to find where each target stops being predictable, which tells the team how precise the measurements need to be. They moved from model tuning to fixing the inputs.
SeismicSoCal · measuring a rare-event system05
A live earthquake detection system on 19 seismometers across Southern California, built during my CBU internship. Before trusting it, I replayed 80 archived days it never trained on and accounted for every alert and every miss.
- 01
Reported every precision next to a chance baseline: shift each declared event by an hour and score it again. Real detections stop matching the catalogue; coincidences in a busy catalogue don't.
- 02
Broke the catch rate down by size (9% at M1–1.5 up to 68% at M3+) and by situation: 83% of isolated M3+ quakes against 60% inside rapid sequences.
- 03
Classified all 58 missed M3+ quakes: 27 came back-to-back with another quake, 16 had fewer than 3 stations with data, 7 were seen but not confirmed and 8 weren't seen. That aimed the fix at swarms.
- 04
Tuned for swarms on validation days only (32 configurations, adoption rule fixed in advance), then scored once on the held-out days: M3+ catch rose from 71% to 78% with no duplicates.
- 05
Checked the home-shaking estimate against what people reported feeling: 94% within one intensity level on 28 held-out quakes.
Academy of Testers · AI checked against official scores06
A full-stack AP/SAT study platform I built as ML & AI Software Developer (Jan 2023 – Jun 2026). Its RAG essay grader is judged against the scores College Board gave.
- 01
Ran an A/B: each typed College Board sample response graded twice, by a baseline with no rubric or retrieval and by the grounded grader. Test essays come from 2023/2024 and the retrieval corpus from 2025, so nothing leaks.
- 02
Used metrics built for ordered scores: mean absolute error, exact and within-one agreement, and quadratic weighted kappa (QWK), which measures agreement beyond chance and counts a miss by two points four times as much as a miss by one.
- 03
Kept APUSH as a held-out check, a subject the grader was never tuned on: error still fell 22%. The samples are small (39 + 8), which I state next to the results, and the config was frozen after the run.
- 04
Diagnosed score compression: nearest-neighbour retrieval returned mostly mid-band essays, so grades drifted to the middle. Stratifying the retrieved pool by score band fixed most of it.
BearLM · measuring retrieval with exact answer keys07
A fully local Q&A assistant over UC Berkeley CS and Data Science courses. Every answer cites the exact PDF page, and I built a benchmark so each retrieval change could be measured.
- 01
Generated the benchmark from real corpus chunks: each question is written from one chunk, so that chunk's ID is the exact answer key. recall@k then asks whether the true source is in the top k results.
- 02
Measured each retrieval stage on it: recall@1 went from 0.56 with dense vectors alone to 0.60 with hybrid keyword search, then 0.82 with a reranker.
- 03
Stated the weak spot: the RAGAS answer-quality runs are small (n = 6) with a local judge, so the retrieval ablation is the stronger evidence.
- 04
Logged usage to SQLite for an analytics page: questions asked, most-cited courses, answered vs. no-source counts, and the course that needs the most help.
NumIsToken · measuring before fixing08
Full-Stack Software Engineering Intern at NumIsToken, Sep 2025 – Jun 2026, on the backend behind token redemptions.
- 01
Profiled the redemption lifecycle APIs before changing anything. The two to three seconds per action came from the same lookups and checks, repeated across a single request.
- 02
Removed the redundant calls and measured again: about 300 ms per action, roughly 85% faster.
- 03
Persisted every step of a redemption behind History and Detail services, so an interrupted one can be found and resumed from its last confirmed state without acting twice on the same payment.
Kigumi Group · root-causing recurring failures09
Backend Software Engineer Intern in Hong Kong, Jun – Sep 2025, on KiguLab, a digital wellbeing and AI-ethics curriculum for ages 8–21.
- 01
Started from AI assistant failures that looked random from outside: the same question would work once and fail the next time.
- 02
Separated them by targeted debugging into two classes with different fixes: unstable backend connections, and malformed queries sent to the GPT and image models.
- 03
Measured coverage before and after (90% to 98%, so the untested code shrank by 80%) and pinned each diagnosed failure with a test.
AperioHub · teaching the analysis habits10
Part-time Research Mentor in AI & ML at AperioHub Advisory, Oct 2026 – present, guiding one high school student through a research project.
- 01
The student picks one of five directions in the first sessions; whichever it is, I teach the habits I use in my own analysis.
- 02
Every claim needs a baseline worth beating, data the model never saw, and the gap reported with its uncertainty. On an open dataset that starts with the cleaning and leakage checks; on a bias question, with a method someone else could re-run.
- 03
Each session reviews what the last task showed, introduces one new idea and assigns one focused task. Recorded sessions and AI-assisted notes keep the project's full record, dead ends included.