I'm Utsa Poddar. I build LLM integrations, retrieval systems, and tooling for AI agents, with deterministic code that checks what a model returns before anything ships. I'm looking for AI Engineer roles in Calgary or remote across Canada.
Physics taught me to distrust a result I can't account for. That habit shapes how I build: models report their uncertainty instead of a single confident number, the deterministic part of a system stays in charge of the decision, and anything I claim about performance is measured rather than assumed.
I hold a BS in Physics from the University of Calgary and am working toward an MS in Data Science at CU Boulder, which includes the university's Artificial Intelligence graduate certificate; both are in progress. Before this I worked as a Business Intelligence Analyst, writing SQL and rebuilding KPI reports, and supervised a team of 11 as an Online Department Supervisor at Canadian Tire. Most of what I build ends up open source: LLM pipelines, retrieval and agent tooling, and estimators you can read and run yourself.
I build systems where the reasoning stays auditable — deterministic components hold the decisions, models advise, and results are measured rather than assumed. Every one is open for inspection.
Python · GitHub Actions · LLM · SEC EDGAR
A weekly investor briefing on GitHub Actions. Ingests SEC Form 4 and 13F filings, corporate crypto treasuries, commodity prices, and 12 news feeds; joins them to a 25-entity watchlist; summarizes through a cross-provider model chain; then emails and publishes each issue every Monday. After four missed issues in a row, I rebuilt the summarizer's retry and fallback path, and every scheduled run since has succeeded.
Python · SQLite FTS5 · Markdown · CLI
Durable, auditable memory for AI coding agents. Markdown notes under a typed schema, full-text retrieval, and a replacement-first correction journal that recovers from a crash mid-write. A recall@5 harness measures retrieval quality instead of assuming it, and credential rejection runs on every write. 71 tests.
Python · AI Agents · Claude Code Hooks
Completion control for AI coding agents. It freezes the objective into prioritized, verifiable requirements, picks the next action from recorded progress, and refuses to report completion until every requirement passes a fresh verifier, so a perceived better idea cannot quietly replace the instruction. The current kernel passes 90 unit tests and 54 structural checks.
TypeScript · Next.js · LLM APIs
In-browser reading guides from PDF and EPUB books. A local extractive engine cites a page or section for every item. An optional provider layer connects the reader's own model, retries transient failures, repairs and validates structured output, and checkpoints each chunk so a model switch resumes instead of restarting. 31 tests.
Python · Streamlit · Supabase · LLM
Scan a barcode or a photo and get a PCOS-aware score with serving advice. The nutrient scoring stays deterministic and rule-based; the model only personalizes and explains that score against your profile. Tested across scoring, vision, barcode parsing, and migrations.
Python · NumPy · SciPy · Bayesian Inference
Predicts how long a graduate course will take you, anchored on courses you have already finished. Course posteriors update from crowdsourced review data under a tempered weight, so heavily-reviewed courses tighten confidence without drowning the prior. Every prediction carries an 80% interval.
Python · scikit-learn · pandas
A reproducible fraud-classification pipeline over 284,807 credit-card transactions. A class-weighted Random Forest reaches 96% precision and 74% recall under stratified evaluation.
JavaScript · Web Audio API · NumPy/SciPy
Eight therapeutic sound modules synthesized live in the browser: AM-modulated 40 Hz gamma, binaural beats, and a user-calibrated notch filter for tinnitus. Every module is gated on peer-reviewed support with its evidence tier and DOI attached, and a NumPy/SciPy script regenerates each spec as reference audio to verify the synthesis independently.
Flutter · Dart · iOS · Android
Cross-platform work timer shipped as downloadable iOS and Android builds, with AltStore distribution for iOS. Automatic phase transitions, boundary alarms, and daily streak tracking keep sessions moving.
University of Colorado Boulder
Graduate coursework in probability theory, statistical inference and estimation, and data science practice — earned an A in each completed course to date. Includes the university's Artificial Intelligence graduate certificate, also in progress.
University of Calgary
Advanced calculus, linear algebra, numerical methods, data analysis, error analysis, statistical reasoning, systems modeling, experimental design
Mimo
Core foundations, OOP, scripting, APIs, AI development, LLM toolbuilding