I build GenAI and machine learning systems that run in production — not just models in notebooks.

I'm currently finishing my Master's in Data Science at Illinois Institute of Technology, and my work is centered around building end-to-end AI systems across:

• Retrieval-augmented generation (RAG)
• Multi-agent LLM systems
• LLM evaluation and quality pipelines
• Recommender systems and ranking
• MLOps, deployment, and observability
• Real-time inference APIs

What I enjoy most is building the full system around the model: retrieval, routing, evaluation, serving, monitoring, and iteration.

I'm especially interested in problems where:

• LLMs need strong retrieval, grounding, and routing
• models are part of a larger production system
• latency, reliability, and monitoring matter
• evaluation is treated as a first-class part of the stack
• engineering decisions matter as much as model choice

LangGraph FastAPI ChromaDB PostgreSQL Redis Prometheus Grafana Kubernetes MCP

• 12-agent LangGraph graph with supervisor routing: deterministic rules first, LLM classifier fallback for low-signal queries
• Hybrid RAG pipeline: ChromaDB vector search + BM25 lexical ranking + Reciprocal Rank Fusion + LLM reranker
• Knowledge graph agent using NetworkX for relationship-style local queries
• Graph-native Human-in-the-Loop (HITL): Streamlit Approve/Reject buttons for web and recency queries
• MCP tool bridge for web search and calculator with automatic local fallback
• TTL caching on Redis with in-memory fallback; dual-layer persistence via LangGraph PostgreSQL checkpointer + conversation store
• Per-user auth, CI/CD via GitHub Actions, Kubernetes manifests, and Prometheus/Grafana monitoring included

View Repository

LangGraph FastAPI OpenAI PostgreSQL GitHub Actions API Prometheus Docker

• Autonomous agent that detects failed GitHub Actions runs, diagnoses the failure, generates a code fix, and opens a PR — without human intervention
• LangGraph state machine with eight explicit stages: ingest → triage → diagnose → reproduce → patch → validate → evaluate → PR or escalate
• Two trigger modes: automatic via GitHub webhook on workflow_run events, and manual via Streamlit operator console
• Patch generation uses OpenAI Responses API with strict JSON-schema output; heuristic fallback for local/demo scenarios
• Every fix must pass lint and tests in a disposable isolated workspace before a PR is opened
• Guardrails enforce allowlisted paths only, block secret-like files, cap patch size, and escalate low-confidence results
• Six Prometheus metrics: success rate, first-patch pass rate, MTTR, escalation rate, false-positive PR rate, retry count

View Repository

LangChain Elasticsearch FAISS BERT MLflow DVC PostgreSQL Docker AWS EKS

• Hybrid RAG pipeline combining BM25 + SentenceTransformer dense embeddings fused via Reciprocal Rank Fusion over Elasticsearch
• distilBERT extractive QA model for direct answer extraction from financial documents
• LangChain agent workflows for context-aware multi-step document reasoning
• Automated evaluation pipeline logging Hit Rate, MRR, and latency per query to MLflow
• PostgreSQL captures user interactions and satisfaction signals for downstream analysis
• Containerized with Docker, deployed via GitHub Actions CI/CD to AWS EKS with full Kubernetes manifests

View Repository

OpenAI Cohen's Kappa Fleiss' Kappa SQLite AWS S3 Streamlit

• Production-grade pipeline for validating annotation consistency and evaluating LLM output quality on QA datasets
• Inter-annotator agreement scoring using both Cohen's Kappa and Fleiss' Kappa across multiple annotators
• Schema validation layer flags malformed or inconsistent annotation records before scoring
• LLM-as-judge scoring via OpenAI API to evaluate response quality at scale
• All results logged to SQLite with AWS S3 for artifact and report storage
• Streamlit dashboard surfaces agreement metrics, judge scores, dataset quality summaries, and evaluation trends

View Repository

QLoRA PEFT Mistral-7B Hugging Face Dual Adapters Instruction Tuning

• Fine-tuned Mistral-7B under resource constraints using QLoRA: 4-bit quantization + LoRA adapters on limited GPU memory
• Two separate adapter sets trained: one for conceptual explanation mode, one for code generation mode
• Runtime routing layer selects the correct adapter based on query intent at inference time
• Instruction-tuned for data science notebook workflows — explanation, debugging, and code generation tasks
• Focus on reliable, consistent outputs for technical use cases rather than benchmark-only performance

View Repository

Multi-Stage Two-Tower Recommender TensorFlow Recommenders FAISS FastAPI MLflow DVC Airflow A/B Testing Prometheus

AutoML LangGraph Assistant LangGraph ChromaDB OpenAI GPT-4o MLflow DVC AWS EC2/S3 MCP Docker

ECG Anomaly Detection — LSTM AutoEncoder TensorFlow Keras LSTM AutoEncoder Unsupervised 97.93% accuracy

Brain Tumor Segmentation PyTorch ResNeXt50-UNet Streamlit LGG MRI

LLM / GenAI LangGraph LangChain LlamaIndex RAG Agentic AI QLoRA PEFT MCP LlamaGuard Vector Search Hybrid Retrieval Reranking

ML / Retrieval / Ranking PyTorch TensorFlow scikit-learn FAISS BM25 XGBoost MLflow DVC

Backend / Infra FastAPI Streamlit PostgreSQL Redis Docker Kubernetes AWS Azure

Monitoring / Ops Prometheus Grafana GitHub Actions

I'm looking for entry-level roles in:

• Machine Learning Engineering
• Applied AI / GenAI
• Data Science with strong ML systems focus

I'm authorized to work in the U.S. on F-1 OPT.

• Email: [email protected]
• LinkedIn: linkedin.com/in/theepankumar
• Portfolio: theepan-portfolio.netlify.app

Outside of work, I like trekking and baking. One clears the head, the other feeds it.