Platform ownership: own the architecture, roadmap and delivery of a platform that turns any uploaded dataset into analysis traceable to its source. Twelve agents over 76 registered tools on Step Functions, 347k lines of TypeScript and Vue across 12 workspace packages, 4,845 tests.
Answer checking: cut wrong answers 85% on 164 published figures by reading every answer back in words and judging it before it is shown. 68% on StatQA, above GPT-4o's best reported 64.83%.
Analyst agent: quadrupled the headline findings a page recovers, 13% to 52-59% on surveys it had never seen, at about $0.05 per 13 surveys. It sees one line per question and never a row; code compiles its choices into plans.
Label reliability: schema-constrained decoding, per-response correlation tokens and citation verification took join integrity to 100% and label churn to zero across 100% of rows. The benchmark pipeline destroyed a label on 1 in 8 rows (12.7%) without flagging one.
EdSpace: built in two days for the EdPlus hackathon, a live office map with room booking, desk claiming, walking routes and a 3D desk per person. Shown to leadership up to the CEO and chairman, and going live for 500+ people.
Rows that silently lost their label, one square per rowFull role→
Pipeline re-architecture: as founding AI/ML engineer, rebuilt a 14-Cloud-Function event pipeline as a single checkpointed LangGraph agent-worker on Cloud Run Jobs.
Scraping at scale: built and operated a Playwright service on Compute Engine that indexes 70,000+ records a day from 650+ locations, rotating proxies and backing off when sites push back.
Entity resolution: Gemini-based dedupe before records reach Firestore and Algolia lifted downstream search dataset accuracy by 25%.
Where a scraped record goes before anyone can search itFull role→
Multi-tenant RAG assistant: led end-to-end development of an assistant used by 1,000+ faculty members to author courses reaching 60,000+ students.
Knowledge graph: combined Neo4j with LLM-generated Cypher queries, making transcript analysis 16× faster by cutting it from 4 hours to 15 minutes.
Retrieval quality: Prompt Flow evaluations for groundedness, relevance and coherence, hybrid vector plus keyword retrieval, and a citation on every answer.
Transcript analysis, 16x faster with a Neo4j graph and generated CypherFull role→
Knowledge Exchange for Resilience, Arizona State University, Phoenix, AZ
Hybrid retrieval: a FAISS and Neo4j layer cut p95 query latency 60% while holding 95% recall.
Latency reduction: Redis caching and SQL query rewrites took API latency from 198ms to 20ms, nearly 90%, verified under simulated peak load.
ETL pipeline: an idempotent pipeline with Postgres constraints and validation checks raised dataset accuracy to 95% across 10,000+ faculty profiles ingested a day.
API latency after Redis caching and SQL rewritesFull role→
Monolith to microservices: split a .NET monolith into microservices on Kubernetes as a strangler migration, cutting resource footprint by 35% and infrastructure cost by 20%.
Release pipeline: SonarQube quality gates that block the merge plus containerized release pipelines, with deploy time down 70% and production incidents down 40%.
Demand prediction: a Redis-backed desk-reservation service for 300+ offices at Fortune 500 companies; reported occupancy improved 30% under COVID-era hot-desk constraints.
Adaptive polling: cut driver-location serving cost 45% with poll intervals keyed to distance remaining, interpolated on the client so movement still looks continuous.
Live order tracking: a Google Maps screen with route polylines and an ETA recomputed on every fix, shipped during a quarter of 25% user growth.
Notification service: WebSocket and Express.js with backpressure-aware fan-out and duplicate-safe reconnection; after launch, reported CSAT rose 28% and support volume fell 35%.
Location polls keyed to distance left, sparse across town, dense near the doorFull role→
Video evaluation harness: scores text-to-video output on physical plausibility separately from semantic fidelity, so impossible motion in a beautiful scene loses points on exactly one axis. Built in PyTorch and Diffusers.
Physics rubric: a frame-sequence rubric for rigid-body motion, with trajectories scored against a constant-gravity analytic baseline rather than by eye.
Controlled comparisons: minimal-pair prompts where exactly one clause changes, with seeds, resolution, frame count and sampler settings held fixed across models.
Scored on two axes, so a pretty clip with impossible motion fails only one, illustrativeFull role→