I build machine learning systems that actually run in production not just models in notebooks.
Right now I’m finishing my Master’s in Data Science at Illinois Institute of Technology, and most of my work sits at the intersection of:
- Retrieval-augmented generation (RAG)
- Multi-agent LLM systems
- Recommender systems and ranking
- MLOps and deployment pipelines
- Real-time inference and monitoring
I enjoy building full systems end-to-end: training, evaluation, deployment, and observability.
Here are a few projects that represent the kind of work I like doing.
I built a multi-agent AutoML system that can generate pipelines, track experiments, and deploy models automatically.
It uses LangGraph for orchestration, MLflow for tracking, and DVC for versioning, with everything containerized and deployed on AWS.
What I cared about here wasn’t just accuracy — it was reproducibility and automation.
This project handles text, images, audio, and video in one retrieval pipeline.
It combines CLIP embeddings, Whisper transcription, vector search, and graph memory with Neo4j.
The interesting part was improving retrieval quality using hybrid ranking and reranking instead of relying on embeddings alone.
I built a hybrid RAG pipeline for financial documents and deployed it to AWS EKS.
The system uses BM25 + dense retrieval, agent workflows, and evaluation dashboards to track retrieval quality and latency.
This project was mostly about infrastructure and evaluation, not just building the model.
A full recommendation pipeline with TensorFlow Recommenders, FAISS retrieval, ranking models, and monitoring.
I focused on:
- Candidate generation vs ranking separation
- Online inference APIs
- Metrics like NDCG and Recall@K
A LangGraph-based system coordinating multiple agents through a FastAPI backend with PostgreSQL persistence.
This was where I spent time designing routing logic, memory, and observability rather than just chaining prompts.
Fine-tuned a Mistral-7B model with dual adapters and routing logic to switch between explanation and code generation modes.
The most interesting challenge here was making a 7B model run reliably on limited GPU memory.
Some additional areas I’ve explored:
- Graph neural networks for fraud detection
- OCR and document AI pipelines
- YOLOv8 optimization with TensorRT
- Healthcare ML (ECG anomaly detection, imaging models)
- Causal inference and experimentation
- ETL and data pipelines
I’m interested in problems where:
- Models are part of a larger system
- Deployment and monitoring matter
- Performance and latency are real constraints
- Engineering decisions matter as much as model choice
Entry-level roles in:
- Machine Learning Engineering
- Applied AI / GenAI
- Data Science (applied / product-focused)
I’m authorized to work in the U.S. on F-1 OPT.
Email: tgandhi1107@gmail.com
LinkedIn: https://www.linkedin.com/in/theepankumar
Portfolio: https://theepan-portfolio.netlify.app
Outside of work, I like trekking and baking. One clears the head, the other feeds it.
