My Work
Here's a collection of projects I've worked on. Each project represents a unique challenge and learning experience.
Featured Projects
Some of my best work and most challenging projects
Verrik — Independent trust layer for AI decisions
2026- Problem:
- In regulated industries, carriers are liable for AI-agent decisions — even vendor models — yet have no independent way to prove a decision was reproducible, fair, and audit-ready before the EU AI Act high-risk deadline.
- Built:
- A "flight recorder and auditor" for AI agents: captures every decision as a durable, tamper-evident trace (SHA-256 hash chain) with deterministic replay, runs four-fifths disparate-impact fairness evaluation in shadow mode, and auto-generates regulator-ready audit reports. Shipped as a TypeScript monorepo with a Postgres-backed Fastify API and a Next.js site.
- Outcome:
- Reproducible, regulator-ready audit trails for autonomous AI decisions — independent of the agent vendor being audited.
Stack
TypeScriptNext.jsFastifyPostgreSQLTurborepoAI Governance
DealLens — Multi-agent AI due diligence
2026- Problem:
- Due diligence means correlating evidence across finance, legal, tech, GTM, ops, and people under heavy time pressure — slow, manual, and easy to miss cross-cutting risk.
- Built:
- A multi-agent platform where a planner agent fans out to six department specialists and a synthesizer cross-correlates evidence-cited findings into a risk-scored investment memo — with cost-aware model routing across providers and real-time SSE streaming.
- Outcome:
- Evidence-cited, risk-scored investment memos generated end-to-end; deployed full-stack on Vercel and Render.
Stack
PythonFastAPILangGraphNext.jsClaudeMulti-Agent
Phishing Shield — Agentic email threat detection
2025-2026- Problem:
- Phishing detection in Gmail/Outlook needs to be real-time, accurate across attack categories, and resilient when offline.
- Built:
- Agentic Chrome extension + FastAPI backend using Claude 3.5 Sonnet for live classification (Phishing / Tech Support Scam / Scareware / Benign). MutationObserver scans incoming mail; local heuristic fallback covers offline mode.
- Outcome:
- Sub-second threat analysis across 50+ test emails with confidence scores and color-coded toasts.
Stack
Chrome ExtensionFastAPIClaude 3.5 SonnetEmail SecurityReal-time
F1 RaceSim — Real-time race strategy with AI
2025- Problem:
- F1 teams reason about pit windows, tire compounds, and weather under heavy time pressure. Existing public sims lack live AI commentary and multi-strategy comparison.
- Built:
- Full-stack simulator (Next.js + FastAPI + Gemini) with multi-strategy comparison, weather-aware predictions, and interactive race visualizations. Deployed on AWS Lambda + Vercel with serverless optimization and AI inference rate-limiting.
- Outcome:
- 40% faster cold-start time post-optimization; real-time AI insights served at race-grade latency.
Stack
Next.jsFastAPIGeminiAWS LambdaVercelPython
Ritematch — AI job matching + résumé tailoring
2023- Problem:
- Candidates send the same résumé to dozens of roles; ATS systems reject what should be relevant matches due to keyword mismatch.
- Built:
- React + Django portal with REST APIs for NLP-based job-résumé similarity scoring and GPT-4o-driven résumé rewriting tuned per JD.
- Outcome:
- 20% improvement in candidate-job match relevance vs. baseline keyword matching.
Stack
DjangoReactGPT-4oRESTLLMsNLP
Verbizz — Real-time business recommendations
2024- Problem:
- Recommendation latency kills conversion; batch pipelines can't react to live user signals.
- Built:
- Cloud-native microservices platform on gRPC + Kafka + Redis + Flask with content-aware recommendation engine. PostgreSQL for warm storage, Kubernetes for orchestration.
- Outcome:
- Scalable AI-first architecture serving real-time personalized recommendations.
Stack
FlaskKafkaRedisgRPCPostgreSQLKubernetesDocker
PitStopAnalytics — F1 Championship Predictor
2024- Problem:
- Predicting F1 championship outcomes requires modeling driver, constructor, and circuit interactions across noisy historical data.
- Built:
- ML pipeline (clustering + PCA + classification) over Ergast F1 API data exploring driver and constructor patterns.
- Outcome:
- Interpretable championship-outcome models with visualized driver/constructor archetypes.
Stack
PythonPandasScikit-learnPCAClusteringJupyter
BizStream — Real-time recommendation engine
2025- Problem:
- Recommendation APIs needed sub-200ms latency at scale; sync data pipelines bottlenecked I/O.
- Built:
- Real-time recommendation system using Kafka + Redis + FastAPI with Docker Compose orchestration and async data pipelines.
- Outcome:
- 35% API latency reduction through smart caching and efficient async I/O.
Stack
KafkaRedisFastAPIDocker ComposePythonAsync
LLMForge — Production sentiment analysis API
2025- Problem:
- Need a deployable sentiment-analysis service with monitoring, framework portability, and stable accuracy.
- Built:
- FastAPI + BERT service with Prometheus monitoring, Dockerized deployment, and PyTorch/TensorFlow portability.
- Outcome:
- 98% accuracy on IMDb; production-ready monitoring and dual-framework support.
Stack
BERTPyTorchTensorFlowFastAPIDockerPrometheusNLP
Formula 1 Insights Dashboard
2025- Problem:
- F1 race analytics scattered across formats; no unified interactive view of driver, team, and race-level data.
- Built:
- Streamlit dashboard with embedded Tableau views for interactive exploration of F1 statistics.
- Outcome:
- Unified real-time view of race data, driver performance, and team analytics.
Stack
StreamlitTableauPythonData Visualization