Featured
Municipal Sustainability AI
Adopted by local governments to benchmark climate investments. 63% MAE reduction on scoring model, 42% retrieval precision gain across 7 prototype generations — 4-stage multi-LLM pipeline with DSPy, BM25 + dense retrieval, and RAGAS gating. UIC CLEAN Initiative.
Customer Churn Analytics Platform
$2.1M in annual revenue at risk quantified — changed the product roadmap. XGBoost + LightGBM at 0.84 ROC-AUC across 15K customer segments, 40+ behavioral features, SHAP explainability. Mirrors the production system built at PlanSource.
Intelligent Operations Command Center
F1 84.9%, FPR 0.15% on anomaly detection — ensemble Isolation Forest + PyTorch autoencoder. LightGBM forecasting at 14.84% MAPE (35.7% over baseline). RAG-powered root cause analysis via Claude API. Airflow + MLflow + FastAPI end-to-end.
All Projects
Document Intelligence Hub
Eliminates the need to manually search documents — ask a question, get a cited answer in seconds. Hybrid BM25 + dense retrieval with Reciprocal Rank Fusion over PDFs, Word docs, and spreadsheets. Runs fully offline, no cloud APIs.
Customer Churn Analytics Platform
End-to-end ML system for churn prediction — mirrors the production system that quantified $2.1M in annual revenue at risk across 15K customer segments at PlanSource. XGBoost with SHAP explainability, served via FastAPI, across five React dashboard views.
A/B Testing & Experimentation Platform
Full-stack experimentation engine that cut analysis time from 3 days to 15 minutes. Supports power analysis, Welch t-tests, SPRT sequential testing, and Holm-Bonferroni correction for multi-variant experiments — built for statistical rigor at scale.
CareSight — Hospital Readmission Prediction
Predicts excess readmission ratios across 18,774 CMS hospital records covering six HRRP conditions. XGBoost pipeline with SHAP waterfall explanations and a three-tab Streamlit dashboard — R² 0.948, F1 0.972 on ERR classification.
QuantifyAI — Quant Alpha Research
Turns raw price data into testable alpha — RSI, MACD, and Bollinger Band signals feed a Random Forest classifier for buy/sell decisions, then Backtrader runs the full portfolio simulation. Live data to backtest result in one pipeline.
Intelligent Operations Command Center
End-to-end AIOps platform: ensemble anomaly detection (Isolation Forest + autoencoder, F1 84.9%), LightGBM capacity forecasting, and RAG-powered root cause analysis via Claude API — orchestrated with Airflow, tracked with MLflow, deployed via Docker.
Municipal Sustainability AI
Progressive AI architecture from single LLMs to agentic workflows, built as an independent study (IDS 596 at UIC). Scrapes 41 climate targets, extracts structured data with DSPy, indexes 714 document chunks for semantic search, and routes queries through specialist LLM roles with self-verification.
FanPulse — Bayesian Marketing Mix Model
Bayesian MMM analyzing 11,972 MLB home games across five seasons — R² 0.81, MAPE 8.3%, $3.2M projected revenue impact. Models geometric adstock, Hill saturation curves, and hierarchical pooling across 30 teams. SLSQP budget optimization surfaced in a six-page Streamlit dashboard.
AI Financial Analyst Agent
Compresses hours of 10-K/10-Q reading into a structured research brief — specialist agents for extraction, ratio analysis, sentiment, and synthesis. RAG over SEC filings with ChromaDB, cited sources in every output.
LLM Observability Platform
Raised retrieval precision 42% across 7 prototype generations by gating every release on 12 RAGAS checkpoints. RAGAS scoring, LLM-as-judge, golden test set versioning, and drift detection — the same eval infrastructure used in the UIC CLEAN research.
dbt Semantic Layer & Analytics Engineering
End-to-end analytics engineering stack on Snowflake — Fivetran ingestion, dbt transformations with 98%+ test coverage, dimensional modeling, and a semantic layer exposing governed metrics to BI tools. Includes CI/CD pipeline with dbt Cloud and GitHub Actions.
SQL Co-Pilot
Ask a plain-English question, get a validated, executable SQL query — handles joins, window functions, and nested subqueries. Schema injection + few-shot selection + self-correction loop cuts hallucinated columns to near zero.
AutoDS — Autonomous Data Science Agent
Drop in a CSV and a target column — get back a tuned model and a research-grade report. Five specialist agents (EDA → feature engineering → model selection → Optuna tuning → report) with full reasoning exposed at every step.
