Building enterprise-grade systems at the intersection of AI/ML and full-stack engineering. I architect production-ready applications that solve real problems — from spatio-temporal graph neural networks to hybrid retrieval-augmented generation systems. Every project I build is designed to scale, deploy, and deliver measurable impact.
- AI/ML Engineering — Graph Neural Networks, RAG pipelines, NLP classifiers, time-series forecasting, MLOps workflows
- Full Stack Development — FastAPI backends, Streamlit frontends, REST APIs, JWT auth, Docker containerization
- Product Engineering — End-to-end ML systems with CI/CD, monitoring, evaluation, and cloud deployment
- Open Source — Actively contributing to reproducible ML tooling and production-grade AI systems
Languages
Frontend
Backend & Databases
Cloud, DevOps & Tooling
ML & Data Science
| Domain | Proficiency | Details |
|---|---|---|
| Graph Neural Networks | Advanced | Spatio-temporal modeling with STGCN, PyTorch Geometric |
| Retrieval-Augmented Generation | Advanced | Hybrid vector + BM25 + graph retrieval, cross-encoder reranking |
| Natural Language Processing | Advanced | TF-IDF, transformers, BERT, sentiment analysis, text classification |
| Computer Vision | Intermediate | EfficientNet, MRI classification, transfer learning |
| Time Series Forecasting | Advanced | LSTM, temporal convolution, Bitcoin direction prediction |
| MLOps & Deployment | Advanced | MLflow tracking, Docker, CI/CD pipelines, Render/AWS deployment |
| Statistical Modeling | Advanced | A/B testing, hypothesis testing, feature engineering |
| Classical ML | Advanced | XGBoost, Random Forest, Logistic Regression, SVM, clustering |
| LLM Agents | Intermediate | LangGraph, tool-use agents, query classification, memory management |
| Vector Databases | Advanced | pgvector, HNSW indexing, cosine similarity, embedding pipelines |
Enterprise Hybrid GraphRAG
Most RAG projects just do vector search. This one combines three retrieval methods — vector search, keyword search, and a knowledge graph — to answer questions that none of those approaches could handle alone.
| Aspect | Detail |
|---|---|
| Stack | FastAPI · pgvector · Neo4j · sentence-transformers · Cross-encoder · Groq (llama-3.1) · Docker Compose |
| Scale | 20+ implemented features · Full ingestion pipeline · Multi-source retrieval |
| Performance | Hybrid RRF fusion · Cross-encoder reranking · Query classification for optimal routing |
| Security | API key authentication · Rate limiting · Query caching · Environment-based secrets |
| Impact | Handles multi-hop entity queries that fail with vector-only or graph-only approaches |
| Repository | Enterprise-Hybrid-GraphRAG |
Upload a PDF → text extraction → semantic chunking → embeddings → PostgreSQL + Neo4j knowledge graph. When you ask a question, a query classifier routes to the best strategy (vector, graph, BM25, or hybrid), merges with Reciprocal Rank Fusion, reranks with cross-encoder, and generates a grounded answer with source citations.
RAG Assistant
Upload any PDF, ask questions, get answers grounded in the actual document. No hallucinations, no guessing.
| Aspect | Detail |
|---|---|
| Stack | FastAPI · pgvector · sentence-transformers · BM25 · Cross-encoder reranker · Groq · Streamlit |
| Scale | Full ingestion pipeline · Evaluation framework (RAGAS) · Conversation memory |
| Performance | Hybrid BM25 + vector search · Reciprocal Rank Fusion · Cross-encoder reranking |
| Security | API key authentication · Query caching · Environment-based secrets |
| Impact | Average evaluation score 0.75/1.0 across faithfulness, relevance, and context precision |
| Repository | Rag |
Three retrieval methods run in parallel: BM25 for exact keyword matches, vector search for semantic meaning, and cross-encoder reranking for precision. Results are merged with RRF and scored across faithfulness, relevance, and context precision.
STGCN Traffic Forecaster
Predicting urban traffic speeds across 207 road sensors using Spatio-Temporal Graph Convolutional Networks.
| Aspect | Detail |
|---|---|
| Stack | PyTorch · FastAPI · Streamlit · NumPy · Pandas · METR-LA Dataset |
| Scale | 207 road sensors · 30-minute forecasting horizon · Interactive live demo |
| Performance | Joint spatio-temporal dependency learning via graph + temporal convolution |
| Security | Containerized deployment (Docker) · Modular project structure |
| Impact | Realistic traffic speed forecasts with interactive slider demonstrating dynamic inputs |
| Repository | stgcn-traffic-forecasting |
Implements STGCN blocks combining graph convolution (propagates features across road network adjacency) and temporal convolution (extracts time-series patterns with gated activation). Streamlit dashboard includes real-time input scrubbing.
Customer Churn Prediction API
End-to-end production ML system with CI/CD, JWT auth, Docker, and cloud deployment.
| Aspect | Detail |
|---|---|
| Stack | Scikit-learn · FastAPI · Streamlit · Docker Compose · GitHub Actions · Render |
| Scale | 7,043 customers · 20 features · ROC-AUC ~0.83 · F1-optimized threshold |
| Performance | Threshold tuning (0.42 vs default 0.50) improved churn recall significantly |
| Security | JWT authentication · Role-based access control · Environment variable secrets |
| Impact | Deployed on Render with automated CI/CD pipeline on every push to main |
| Repository | churn-prediction-ml-api |
Scikit-learn Pipeline handles imputation, encoding, scaling, and Logistic Regression in a single serialized object. GitHub Actions workflow runs tests, builds Docker image, and publishes to Docker Hub automatically.
Task Planner Agent
AI agent that breaks high-level goals into tasks, plans execution order, and runs them with real tools — pausing for human input when needed.
| Aspect | Detail |
|---|---|
| Stack | LangGraph · LangChain · Groq (llama-3.3-70b) · FastAPI · Docker |
| Scale | Multi-node state machine · Plugin-style tool registry · Session memory |
| Performance | LLM-based planning + review loop · Retry on failure · Dependency-aware task ordering |
| Security | Environment-based API keys · Docker containerization |
| Impact | Interactive 3-option clarification UI for ambiguous decisions mid-execution |
| Repository | -Task-Planner-Agent |
LangGraph orchestrates planner → executor → reviewer → router nodes. When the agent hits ambiguity, it shows a clean 3-option card instead of guessing. After each task, an LLM reviews completion and retries if needed.
Ticket Categorizer
Classify support tickets into Billing, Technical, HR, or General using NLP — with confidence-based human review routing.
| Aspect | Detail |
|---|---|
| Stack | TF-IDF · Logistic Regression · Streamlit · HuggingFace Datasets |
| Scale | ~12,000 English tickets · 4 categories · 3 model comparison |
| Performance | 89% accuracy · Macro F1 0.70 · Confidence-based auto-routing |
| Security | Human-review fallback for low-confidence predictions |
| Impact | Auto-routes high-confidence tickets, flags ambiguous ones for human review |
| Repository | Ticket-Categorizer |
Compares Naive Bayes, Logistic Regression, and Linear SVM. Logistic Regression chosen over SVM (94.6% accuracy) because it supports predict_proba for confidence-based routing. Human-review triggers on top confidence < 50%, top-2 margin < 15%, or < 4 meaningful words.
Brain Tumor MRI Classifier
Brain Tumor MRI Classifier using EfficientNetB0 — classifies Glioma, Meningioma, Pituitary & No Tumor.
| Aspect | Detail |
|---|---|
| Stack | TensorFlow · EfficientNetB0 · Transfer Learning · OpenCV |
| Scale | 4-class classification · Transfer learning from ImageNet weights |
| Performance | 91% accuracy on test set |
| Security | Model serialization · Reproducible training pipeline |
| Impact | Medical imaging classification with production-grade accuracy |
| Repository | brain_tumor |
Fine-tuned EfficientNetB0 with transfer learning for multi-class MRI classification. Achieves 91% accuracy across 4 tumor types using data augmentation and class balancing.
Personal & University Projects
Designed and deployed 6+ end-to-end ML systems including GNN-based traffic forecasting, hybrid RAG pipelines with knowledge graphs, MLOps workflows with CI/CD, and LLM-powered task planning agents.
PyTorch FastAPI Docker MLflow Neo4j LangGraph GitHub Actions
| Recognition | Details |
|---|---|
| Enterprise GraphRAG | Built hybrid vector + BM25 + knowledge graph RAG system with 20+ features |
| STGCN Traffic Forecasting | Implemented spatio-temporal GNN for 207-sensor traffic prediction |
| MLOps Pipeline | Automated CI/CD with GitHub Actions → Docker Hub → Render deployment |
| RAG Evaluation | Achieved 0.75/1.0 average across faithfulness, relevance, context precision |
| 6+ Production Projects | Full-stack ML systems with auth, testing, containerization, and cloud deployment |
| Open Source Contributions | Actively building reproducible ML tooling and documentation |
Learning:
- Advanced MLOps patterns and model monitoring
- System design for ML infrastructure
- Distributed training and inference optimization
Building:
- Enterprise-grade RAG systems with knowledge graphs
- Production ML pipelines with CI/CD automation
- LLM-powered agents with tool use and memory
Exploring:
- Graph Neural Networks for complex systems
- Multi-agent architectures
- Edge deployment for ML models
Open To:
- DS / ML campus placements 2026
- Open source collaborations
- AI/ML engineering internships
"Every dataset has a hidden story waiting to be uncovered."