A model is the easy half. What decides whether it survives contact with reality is everything to its left and right — where the data comes from, whether you can trust it, and what metric you actually optimise.
Live in production
The same prescription can cost 3× more one block away.
CheaperRx makes that spread visible — search a drug, see every pharmacy in your city ranked by real price, and get told when a generic equivalent exists. Auth, subscriptions, crowdsourced data and an admin verification queue, running today.
Authorisation lives in the database
Row-level security on all eight tables. A bug in a React component cannot leak another user's medication list — Postgres refuses the row.
Stripe webhooks are the only truth
The client never sets role = 'pro'. Checkout success is a UI event; the webhook is what grants entitlement.
Current state vs. ledger
A unique index guarantees one live price per drug–pharmacy pair, so the hot read path never deduplicates. History is appended separately.
Next.js 14TypeScriptPostgreSQL
SupabaseStripeGIN full-text search
scikit-learn · imbalanced-learn
Space Object Classification
Debris massively outnumbers rocket bodies, so a model that always guesses "debris" scores well and is useless. Class imbalance was the whole problem. The jump from linear SVM to RBF says the boundary in orbital-element space is genuinely non-linear.
97.4% Random Forest
SMOTE to 8,431/class
Per-class metrics
TensorFlow · Keras · VGG19
Pediatric Pneumonia Detection
Triage support for chest radiographs, where a false negative costs far more than a false alarm. VGG19 transfer learning with augmentation and regularisation tuned for a small, imbalanced clinical dataset — recall is the objective, not accuracy.
~88% accuracy
Augmentation + BatchNorm
EarlyStopping
Python · pandas · Power BI
Alberta Oil Sands Production Analysis
Alberta's regulator publishes ST39 as a document, not a dataset: months run across the page, commodities stack down it. Unpivoting three annual workbooks into one tidy fact table — 9 operators, 36 months, 251 monthly series — is what turns every downstream measure into a one-line aggregation.
Gibson +15.5% (2022–24)
May output 13% below mean
12,084-row fact table
Python · pandas · Streamlit
HR Analytics Automation Pipeline
Five inconsistent workbooks, twelve hours of manual consolidation, a report nobody fully trusted. The fix was not a better spreadsheet — it was a 12-rule validation gate that fails loudly instead of quietly corrupting the numbers downstream.
12h → 90s
12 validation rules
Self-service upload
