- Start here — FraudWatch Week 1
- Which domain should you pick?
- Finance flagships
- Energy flagships
- How they map to your 6–8 month skills
- Status tracker
- Stack & data sources
Start here — FraudWatch Week 1
Locked primary: FraudWatch (then FilingPilot later). Follow these steps
in order. Code lives in the repo folder fraudwatch/.
Step 0 Setup venv + pip install -r fraudwatch/requirements.txt Step 1 Write the problem in 5 lines (who, costs, why accuracy fails) Step 2 Run: cd fraudwatch && python 01_baseline.py Step 3 Read PR-AUC table + artifacts/pr_curve_week1.png Step 4 Build-log note: which model wins and why Step 5 Drill: precision vs recall for fraud queues
Full walkthrough: fraudwatch/README.md. First script: 01_baseline.py.
When Step 2 works on your machine, tell me — we’ll add 02_thresholds.py
(operating-point / cost matrix) next.
Which domain should you pick?
Want GenAI / LLM interviews? → Finance docs RAG (+ fraud/credit baseline) Want MLOps / time-series signal? → Energy load forecasting + anomaly detection Want classic ML depth? → Credit risk or fraud (imbalance, thresholds) Unsure? → Start with FraudWatch (fast feedback), add FilingPilot later
Default pick if you want the broadest interview story: FraudWatch (tabular ML) → then FilingPilot (RAG over financial documents). Same “risk & trust” narrative, two complementary demos.
Finance flagships
1. FraudWatch — transaction fraud detection Recommended start
Problem: flag fraudulent payments without drowning analysts in false alarms.
Why interviews love it: class imbalance, precision/recall tradeoffs, threshold tuning, feature leakage traps, cost-sensitive decisions — all high-frequency ML interview topics with a clear business story.
- Data: public credit-card fraud sets (e.g. Kaggle creditcard / PaySim-style)
- v1: logistic + random forest / LightGBM baselines; PR-AUC + precision@k
- v2: time-aware split, calibration, cost matrix (false alarm vs missed fraud)
- v3: FastAPI scoring endpoint + simple analyst dashboard
- Stretch: LLM that explains why a score fired (using feature contributions)
Interview line: “I optimised for precision at a fixed recall so the review queue stayed usable — here’s the threshold curve.”
2. CreditRisk — default / risk scoring
Problem: predict probability of default; set an approve/reject threshold.
- Data: Lending Club–style or UCI credit datasets
- Skills: logistic regression as regulator-friendly baseline, boosting, ROC vs business cutoffs, fairness checks (optional but impressive)
- Stretch: model card + adverse-action style explanation text
Slightly more “regulated industry” flavour than fraud — great for fintech / banking roles.
3. FilingPilot — RAG over financial filings / research GenAI signal
Problem: ask questions over 10-Ks, earnings transcripts, or investor PDFs and get answers with citations (no hallucinated numbers).
- Data: SEC EDGAR filings (public), or a small hand-picked PDF pack
- v1: chunk → embed → retrieve → generate with source links
- v2: 20-question eval set (numeric facts + policy questions); measure retrieval hit rate + citation accuracy
- v3: hybrid search (keyword + vectors), refusal when evidence is weak
Interview line: “RAG vs fine-tune: filings change quarterly, so I kept knowledge in the index and versioned the prompt — here’s my eval harness.”
4. MarketSense — light forecasting (optional)
Volatility or direction models on public price data. Use carefully: easy to fool yourself with leakage. Only take this if you can defend time-based splits and walk-forward validation in an interview. Prefer FraudWatch/CreditRisk if unsure.
Energy flagships
5. LoadCast — electricity demand forecasting Strong MLOps story
Problem: forecast short-term load (hourly/daily) so operators can plan.
- Data: public ISO / utility load + weather (e.g. regional open datasets)
- v1: strong baselines (seasonal naive, linear, gradient boosting on lags/calendar/weather)
- v2: walk-forward validation, MAPE/RMSE by hour-of-day, residual plots
- v3: scheduled retrain job + drift monitor (input weather/load shifts)
- Stretch: LLM ops brief that summarises “what changed this week” from metrics
Interview line: “Forecasting isn’t accuracy alone — I report error by peak hours and monitor drift after weather regime changes.”
6. GridPulse — sensor / equipment anomaly detection
Problem: spot abnormal behaviour in SCADA-like time series before failure.
- Data: public predictive-maintenance / turbofan / sensor datasets
- Skills: unsupervised / semi-supervised anomaly scores, alert fatigue, precision of alarms, human-in-the-loop review queue
- Stretch: RAG over maintenance manuals when an alert fires
7. PowerPrice — energy price forecasting (advanced)
Day-ahead price prediction. High domain credibility for energy trading / flexibility roles; harder data cleaning. Treat as a second project after LoadCast.
How they map to your 6–8 month skills
These are domain projects — they don’t teach the curriculum, they consume the skills as you learn them:
Status tracker
We’ll lock a primary + secondary after you choose. Current proposal:
Stack & data sources
Classic ML scikit-learn, LightGBM/XGBoost, pandas Deep learning PyTorch (when you need sequences / custom nets) GenAI / RAG embeddings + Chroma/FAISS, then an LLM API API / ship FastAPI + simple UI; monitor latency, scores, drift Finance data Kaggle fraud sets, Lending-style credit, SEC EDGAR Energy data public ISO load + weather; NASA/C-MAPSS-style maintenance sets
What “done” looks like for interviews
- Problem statement + business metric (not just accuracy)
- Leakage-safe validation (time-based when needed)
- Baseline vs best model table
- Failure analysis (when it breaks)
- Small demo (notebook → API)
- One-page system design on this site