I scope, plan, and ship projects — end‑to‑end.

▸ Problem

Supply chain teams react to disruptions after they surface in weekly reports. News, commodity pricing, and shipping signals live in separate tools — nobody correlates them in real time.

▸ Approach

• Ingested 4 external feeds (GDELT news, commodity APIs, AIS shipping, FRED macro) into a DuckDB warehouse
• Ran hourly Python jobs to cluster related events by topic + region
• Routed clusters through Claude for summarization + risk scoring (0–100)
• Surfaced ranked feed in a Streamlit dashboard with supplier/region drill-down

▸ Outcome

• Designed to surface disruption signals 12–36h ahead of top-tier reporting cycles
• Processes ~3,500 events/day at under $4/month operational cost
• LLM cost optimized with Claude Haiku for scoring, Sonnet for only top-ranked clusters

▸ What I learned

Cost-routing between LLM tiers matters more than prompt engineering at scale. 90% of quality comes from clean retrieval.

▸ Problem

UN Comtrade raw data is messy, denormalized, and nearly unusable for ad-hoc questions like "who depends on Russian wheat?". Analysts burn hours on SQL plumbing instead of analysis.

▸ Approach

• Modeled trade flows into a star schema (fact_flows, dim_country, dim_commodity, dim_year)
• Built reusable SQL views with CTEs + window functions for growth, rank, and share metrics
• Implemented Herfindahl-Hirschman (HHI) indices to quantify concentration risk
• Layered Plotly drill-downs over Streamlit filters for ad-hoc exploration

▸ Outcome

• 116K rows queried in under 150ms via DuckDB
• 3 reusable analysis templates: concentration, growth, bilateral dependency
• Loads in < 2s cold; interactive drill-down without rebuilds

▸ What I learned

A well-modeled 100K-row warehouse beats an unmodeled 10M-row lake every time. Star schemas are still the right answer.

▸ Problem

Students spray-apply to roles without understanding fit. Generic CVs get filtered. Tailoring manually for each listing takes 30+ min.

▸ Approach

• Scraping: Playwright microservice on Render for JS-rendered listings (bypassing Vercel serverless limits)
• Scoring: Claude Sonnet for reasoning-heavy match analysis
• Generation: Gemini Flash for cheap CV/cover-letter rewrites
• Frontend: Next.js + Tailwind, deployed on Vercel

▸ Outcome

• End-to-end match + tailored doc pair in ~20s
• Operational cost under $3/month across both APIs
• Match score correlates with call-back rate in personal testing (n=40)

▸ What I learned

Split the pipeline by cognitive difficulty: expensive models for reasoning, cheap models for formatting. You don't need a bigger model, you need better routing.