Filter by area, or tap any card for the details.
🏆 Winner · LPL Financial Hackathon 2026
AI research agent for financial advisors, deployed on AWS.
Next.js · AWS Lambda · DynamoDB · GitHub Actions
- My role: backend and cloud engineer on a team of 4. Teammates built the multi-model chat, agent runtime and terminal UI.
- Owned the AWS backend architecture, deployment, data layer and reliability.
- Built a cross-device chat-history REST API on a DynamoDB single-table design, and distributed rate limiting with DynamoDB counters across Lambda instances.
- Added an error classifier with retry and backup-model failover, plus per-model CloudWatch metrics.
- Built a read-only MCP server and GitHub Actions CI/CD with OIDC, with zero stored credentials.
- Passed 17/17 smoke checks on the first full deploy across 5 Bedrock models.
Hackathon winner, team of 4
🏆 1st Place · AI Track, ALI Builds Hackathon (UChicago)
No-account healthcare navigator for uninsured patients, in 15 languages.
Next.js · TypeScript · Claude API · Vercel
- Helps uninsured patients find care without making an account, in 15 languages.
- Two-phase Claude vision pipeline that extracts details from documents and reconciles them across documents.
- Pulls live data from 5 public health and drug-pricing sources (HRSA, CMS, NADAC, RxNorm, openFDA).
- Built with a team of 4.
Hackathon winner, team of 4
🥈 2nd Place · FitchGroup Codeathon 2025
Emissions Prediction ML System
Quantile-ensemble model that predicts companies' Scope 1 and 2 emissions with calibrated uncertainty.
Python · CatBoost · LightGBM · Quantile Regression · Calibration
- Achieved R² = 0.65 (Scope 1) and 0.48 (Scope 2) with leakage-safe target encoding and quantile ensembles.
- Isotonic calibration on top of the ensemble for well-calibrated uncertainty estimates.
- Modeled $174M risk reduction for portfolio-level emissions estimation.
2nd place, FitchGroup Codeathon 2025, team
Local-first lecture app with GPU transcription and a RAG tutor that cites the exact lecture and timestamp.
Python · FastAPI · React · TypeScript · SQLite · Whisper
- Records or imports lectures, transcribes them on a local GPU with Whisper, and turns them into timestamp-linked study notes.
- Ingests lecture slides (PDF and PPTX) as source material.
- The RAG tutor numbers its sources before the model sees them, so it can't invent citations.
- Swappable local (Ollama) or cloud (Claude, Gemini, OpenAI-compatible) models for each task, with a background worker queue.
Core app done, quizzes next
Resume-to-job matcher using semantic matching and LLM feedback.
Python · FastAPI · spaCy · SBERT · Gemini API · Redis · Docker
- I built it in my sophomore year after two months of applying and hearing nothing back. It showed me which skills kept coming up in the jobs I wanted, so I built projects around them and landed my Mooremac internship.
- Three-stage pipeline: spaCy extraction, SBERT semantic matching and Gemini recommendations.
- Scores a resume in under 3 seconds, with ~85% agreement with manual review.
- Redis caching and async processing cut latency ~3x.
Being rebuilt as a Chrome extension
Prediction-Market Trading Bot
No public repo
Forecasting and risk work on an open-source bot for Chicago temperature markets.
Node.js
- Validated the forecast model against 730 days of NOAA records and 16 months of hourly airport data, cutting top-pick overconfidence from 6.4 to 0.9 percentage points.
- Fixed a bug that mispriced ~23% of days.
- Replayed 550+ stored market snapshots to trace losses to 3 root causes.
- Added risk controls: an exposure cap, a stop-loss and take-profit monitor, and fee-aware edge math.
- Grew the test suite from 72 to 104 tests.
Ongoing
LLM Training on a Supercomputer
Trained an LLM from scratch and fine-tuned a speech model on NERSC's Perlmutter.
Python · PyTorch · Slurm
- Trained a character-level LLM from scratch, then fine-tuned a speech model on self-built datasets in a 4-GPU, 50-epoch run via Slurm.
- Diffed checkpoints to explain why the model sounded unchanged: the pruned weights moved ~40x less than the raw ones.
- Presented a FastAPI demo to the cohort.
Argonne ALCF Intro to HPC Bootcamp, Aug 2026
Campus Route Planner
Code private
Shortest-path routing over UIC's OpenStreetMap data.
C++
- Built a weighted directed graph over UIC's OpenStreetMap data with 7,386 vertices and 22,444 edges.
- Leak-free Dijkstra with a min-heap, building-to-footway linking and JSON parsing.
Done
Responsible-AI recommender over public Census, migration and rent data.
Python · pandas · scikit-learn · MCP
In progress, Break Through Tech AI Studio
Misinformation Classifier
Compares logistic regression, fine-tuned BERT and LLMs on six-class truthfulness labels.
Python
In progress, CS 412, team of 3