Credit Decisioning: Verified Adverse-Action Notices
When credit is declined, the applicant must be told the specific reasons. A calibrated model scores the application, code writes the notice from a reviewed template, and an independent eight-check verifier, proven with 14 kinds of mutation tests, confirms it before a human decides. A read-only AI analyst explains cases, and code checks every number it states. Built on a public benchmark, not a real lender.
Cohort Risk & Review Lab: Fair, Clinician-Reviewed Risk Stratification
Care-management teams can only call so many patients, and a risk score that looks accurate on average can still under-serve one group. This research prototype, built on synthetic patient data, pairs a calibrated risk model with per-subgroup fairness audits and an AI agent that assembles the evidence, while a clinician makes every decision. It also reproduces how training on cost instead of illness burden skews who gets enrolled.
Trade Surveillance Agent: Insider-Trading Review That Shows Its Work
An AI investigator for SEC Form 4 insider filings that gathers evidence, drafts a conclusion, and checks every claim against what it actually retrieved before publishing a clear, flag, or escalate outcome for a human reviewer. Process audits measure how the agent worked, and the project openly reports where accuracy still falls short. A research prototype on public filings and synthetic firm policy.
Clinical Care Navigator: A Patient-Facing Assistant That Is Allowed to Refuse
An assistant that explains a patient's own record is useful; one that answers everything is a liability. This synthetic-data prototype puts independent safety checks before and after the AI, so dosing questions are refused, emergencies get urgent guidance, and answers that conflict with the record, such as a critical lab value, are overridden and sent to a clinician. Safety is built into the workflow, not left to a prompt.
RxFlow Pharmacy Assistant: AI-Powered Healthcare Workflow Automation
A sophisticated conversational AI system designed to streamline pharmacy workflows, automate prescription refills, and enhance patient care through intelligent escalation handling and multi-turn dialogue management. Features advanced state machine architecture, healthcare system integrations, and enterprise-grade performance optimization achieving 75% reduction in processing time and 90% first-contact resolution rate.
Claim Processing Agent: Agentic RAG for Auto Insurance
A production-grade Agentic RAG system that fully automates the evaluation of auto insurance claims. Combines semantic policy retrieval, CSV-backed validation, real-time web search fraud detection, and dual LangChain/LangGraph and Smolagents pipelines to deliver structured, auditable coverage decisions — with 58 automated tests, FastAPI backend, Streamlit UI, and Docker deployment.
DualLens Analytics: RAG-Powered Investment Analysis
A production-grade RAG-powered investment analysis platform combining quantitative financial metrics from Yahoo Finance with qualitative AI-initiative insights extracted from company PDF reports. Features LangChain orchestration, ChromaDB vector store, LLM-as-Judge evaluation, and an interactive Streamlit dashboard with 36+ automated tests and Docker-ready deployment.
GenAI Docs Helper: Intelligent Document Assistant
A production-ready agentic RAG system leveraging LangGraph state machines and advanced AI workflows for intelligent document retrieval. Features sophisticated hallucination detection, multi-strategy retrieval with confidence scoring, and enterprise-grade performance optimization achieving 92% retrieval accuracy and 98% factual accuracy in responses.
Forecasting Engine for Retail Analytics
Developed a highly scalable and accurate demand forecasting engine utilizing Ray distributed computing framework and Vertex AI on Google Cloud Platform. This enterprise-grade solution handles complex retail scenarios across multiple product categories and promotional strategies, featuring advanced machine learning workflows, BigQuery integration for extensive preprocessing, and parallel processing capabilities that deliver exceptional forecasting accuracy and operational efficiency.
Car Insurance Telematics Risk Assessment
Developed a comprehensive machine learning system to predict insurance claim probability and severity using telematics data.
Features include advanced feature engineering, XGBoost modeling with Optuna hyperparameter tuning, and a modular inference pipeline.
GenAI Job Finder: AI-Powered Career Discovery Platform
A sophisticated conversational AI platform leveraging LangChain and LangGraph state machines to create intelligent career guidance experiences. Features multi-LLM architecture (GPT-3.5, Llama 3.2), advanced prompt engineering, and an interactive AI assistant that provides real-time personalized career advice through natural language conversations.
PipeLearner: Machine Learning Optimization Tool
Created PipeLearner, a Spark distributed package that cuts down model training and deployment time by over 50%, streamlining the process and enhancing productivity across data teams.
Optimization of Promotional Forecasting Platforms
Transformed the promotional forecasting platform at a retail company to a PySpark distributed computing framework, revising logics and models in AWS. This led to a 2% reduction in error and a significant cut in runtime by nearly 70%.
Warehouses Orders Forecasting Challenge
Developed a solution for predicting grocery delivery orders at Rohlik warehouses across Europe, enhancing workforce allocation, delivery logistics, and inventory management through accurate forecasting.
MMA Fighter Data Analysis and Ranking
Conducted a comprehensive analysis of MMA fighters using data scraping, ranking, and visualization techniques. This project provides insights into fighter performance and serves as a tool for fans and analysts.
