Expense Claims System
A role-based expense workflow from receipt upload through approval and finance reporting.
Next.jsTailwind CSSFastAPIPostgreSQLSQLAlchemyAWS S3Gemini Vision
Built the receipt-to-payment workflow across the frontend, API, database, and storage layers. Backend rules cover role and team access, claim-state transitions, manager self-approval prevention, duplicate-receipt warnings, duplicate-payment blocking, monthly limits, and audit events. Receipt extraction creates an editable draft; payment is simulated and demo authentication is not production-grade.
ExplainOps — AI Incident Triage
Three simulated event streams become an explainable, ranked incident queue.
FastAPIScikit-learnLangChainGroqSQLiteGitHub Actions
Fuses simulated ticket, log, and chat inputs; ranks incidents with a Logistic Regression model and a deterministic weighted fallback. Groq/LangChain generates grounded explanations for the top five, while false-positive feedback adjusts per-service bias and model weighting. The end-to-end pipeline is tested and runs in GitHub Actions.
ZENO — Valorant Stats & Guides
A server-rendered gaming product with player lookup, stats pages, leaderboards, and guides.
Node.jsExpress.jsMongoDBMongooseEJSBootstrapAxios
Built Express routes and EJS views for player profiles and match history, leaderboards, maps, weapons, agent guides, and pro-player settings. MongoDB/Mongoose models back application data, while Axios integrates external Valorant data. Authentication code is basic configuration and middleware scaffolding, not a hardened auth system.
RAG Document Q&A & Retrieval Evaluation
Document-grounded answers with page citations and a reproducible retrieval evaluation.
FastAPIQdrantBM25 + RRFCross-encoderSentence TransformersGroqLangSmithStreamlit
Implemented dense and BM25 hybrid retrieval with Reciprocal Rank Fusion, cross-encoder reranking, page-level citations, and a grounded not-found response. On the repository's 29 answerable golden-set questions, the full pipeline reports Recall@4 1.000, MRR 1.000, and NDCG@4 1.000 at about 298 ms evaluation latency; this is a small test set, not a production-scale benchmark.
Real-Time Multilingual Call Resolution
Transcribes support calls and routes them with explicit, deterministic rules.
FastAPILangGraphWhisperspaCyGroqStreamlit
A four-node LangGraph workflow uses local Whisper transcription, spaCy language detection, keyword-based intent recognition, and deterministic support routing. Groq generates audit explanations but does not choose the route. FastAPI exposes audio processing, audit-history, and event-replay endpoints, with a Streamlit dashboard for review.
Stroke Risk Prediction API classical ML
A trained classifier served through a documented prediction API.
PythonScikit-learnPandasNumPyFastAPIDocker
Trained a stroke-risk classifier with preprocessing, feature engineering, and class-imbalance handling on a public patient dataset. FastAPI exposes a POST /predict endpoint and the repository includes a Dockerfile for local containerized execution. The model is a demonstration, not a clinical diagnosis or medical decision tool.
