Physics-trained developer building data pipelines, applied ML, and LLM-powered systems. I like problems where a model has to survive contact with messy real-world data — and where the reasoning is auditable rather than a black box.

Portfolio · LinkedIn · utsapoddarjobs@gmail.com

• Data & ML — Bayesian modeling, uncertainty-aware estimation, classification pipelines
• LLM Systems — treating model output as advisory and keeping a deterministic component authoritative
• Automation & Backend — scheduled pipelines, multi-source ingestion, APIs, CI

Alpha Digest — A weekly investor-intelligence pipeline that runs itself. Ingests SEC EDGAR Form 4 and 13F filings, corporate crypto treasuries, commodity prices, and 12 financial RSS feeds; cross-references them against a 25-entity watchlist; summarizes with an LLM behind a three-model retry chain; renders with Jinja2; then emails subscribers and publishes to GitHub Pages — every Monday, on GitHub Actions cron. Read the current issue →

MSDS Time Estimator — Bayesian estimator predicting how long a CU Boulder MSDS course will take you, anchored on courses you've already finished. Course posteriors update from crowdsourced review data under a tempered weight so heavily-reviewed courses tighten confidence without drowning the prior; outputs are uncertainty-aware with 80% intervals. Full derivation in MODEL.md.

PCOS Food Scanner — Scan a barcode or a photo, get a PCOS-aware score and serving advice. Nutrient scoring is deterministic and rule-based; the LLM only personalizes and explains it against your profile. Streamlit + Supabase, with tests over scoring, vision, barcode parsing, and migrations.

Engram — Durable, auditable memory for AI coding agents. Markdown notes under a typed schema, SQLite FTS5 retrieval, a replacement-first correction journal that recovers from a crash mid-write, and a recall@5 harness so retrieval quality is measured rather than assumed. Credential rejection and reasoning-stripping run on every write. 71 tests.

DopaKernel — A control kernel for AI coding agents that keeps the authorized objective as a hard gate rather than a weighted score, so a perceived "better idea" can't silently replace the instruction. Separates adherence, verified progress, and completion evidence.

Cadence — Eight therapeutic sound modules synthesized live in the browser with the Web Audio API: AM-modulated 40 Hz gamma, binaural beats, and a user-calibrated notch filter for tinnitus. Every module is gated on peer-reviewed clinical support with an evidence tier and DOI as structured fields, and the exclusions are documented as deliberately as the inclusions. A NumPy/SciPy script regenerates each spec as reference WAVs to check the JS synthesis against an independent implementation. Try it →

Sift Work Timer — Cross-platform Flutter work timer (iOS / Android / macOS / Windows) with alarms, phase transitions, and streak tracking. Distributed via AltStore and Google Play.

Languages Python · SQL · JavaScript · Dart · TypeScript · MATLAB Data & ML pandas · NumPy · SciPy · PyMC · scikit-learn · Jupyter Backend & Infra FastAPI · Streamlit · SQLite/FTS5 · Supabase · GitHub Actions LLM Gemini · Groq · NVIDIA NIM · prompt/tool design · retrieval

LiDAR room designer (prototype) · Physics simulations · SQL analytics · n8n workflows · Fraud detection · Fluid dynamics

I'm open to data science, ML, and backend/automation roles. Based in Calgary, Alberta. If something here is relevant to your team, reach out at utsapoddarjobs@gmail.com.