Software Engineer, Applied AI @ Teclusion AI
- Co-built Visentix, a multi-tenant privacy-intelligence platform, as one of two engineers — ingestion to reporting.
- Kept the AI out of the arithmetic: confined the LLM to clause classification while 14 versioned deterministic formulas compute every score, with SHA-256 reproducible reports.
- Built 384-dimensional MiniLM semantic retrieval in pgvector against a live peer benchmark.
- Built an evaluation harness over a 200-sample human-labelled gold set with confidence-based suppression.
- Shipped a document RAG service over Google Drive, SharePoint and S3 with tenant-filtered retrieval and cited answers.
- Mined a regulatory corpus through nine scraping connectors across HHS OCR, the FTC, SEC EDGAR and the CPPA.
- Isolated tenants at a single query chokepoint, blocked SSRF, and eliminated idle GPU spend with scale-to-zero inference — on a self-hosted Qwen3-8B with row-level security across 56 tables and a 1,000+ test suite.
AI Engineering Intern @ Teclusion AI
- Shipped a dynamic pricing service blending a rule engine with model output, served through an ONNX model registry.
- Built a booking assistant as an eleven-state deterministic state machine that confines the LLM to slot extraction.
AI Intern, Emerging Tech @ PricewaterhouseCoopers
- Built a full-stack agentic financial-advisory platform on FastAPI, Next.js and Firebase, grounding Gemini in a custom knowledge graph — median response under 2s.
- Co-authored a white paper on AI-enabled software development, researching LangChain and agentic architectures.
A credit-decisioning platform for applicants banks can't score — underwriting from verified bank-transaction evidence (Account Aggregator) instead of a bureau score. A 9-gate deterministic policy engine makes every call with zero AI in the decision path, an 11-feature cash-flow scorecard (ROC-AUC 0.787, KS 0.45) explains each contribution, and a hash-chained audit ledger keeps it all auditable.
A reproducible study benchmarking five model families — ARIMA, GARCH, Prophet, KNN, and feed-forward neural nets — for 30-day stock-price forecasting, weighing accuracy (RMSE, MAPE) against complexity. Every comparison is grounded in statistical validation: ADF stationarity tests, ACF/PACF analysis, Ljung-Box diagnostics, and GARCH volatility modelling for risk-aware forecasts.
A dynamic pricing engine that pairs ARIMA/SARIMA time-series forecasting with ML to predict demand and price elasticity for every SKU. It reacts to competitor moves within business-set margin and brand-price floors — driving a 6–10% lift in optimized revenue and cutting manual pricing effort by 50%.
A serverless crisis-comms PWA spanning three transport tiers (WebRTC over Wi-Fi, animated QR sync, printed posters), deployable by URL in under 30 seconds with zero backend. A gossip sync engine with Bloom-filter anti-entropy, Ed25519 signatures, and NaCl end-to-end encryption hits sub-200 ms LAN sync — all in a 180 KB gzipped bundle.