My research focuses on three-dimensional mirror symmetry — classifying dual pairs in supersymmetric quantum field theories with minimal supersymmetry, an open problem for three decades. My doctoral work developed a systematic framework for constructing these dual pairs, with findings published in Physical Review D and the Journal of High Energy Physics.
📄 Find my publications here
Alongside the PhD, I have been building end-to-end ML engineering skills: agentic workflows, retrieval systems, geometric deep learning, time series forecasting, and domain-adaptive LLM training. The repositories below are the output of that work.
Languages
Data Handling
ML & Deep Learning
LLMs & Agentic AI
MLOps & Deployment
| Project | What it does | Stack |
|---|---|---|
| MolPredict Inference API · API Service | A GATv2 inference API service for predicting HOMO-LUMO gaps of organic molecules; inference tools additionally deployed as an independent MCP server over SSE. | PyTorch Geometric, FastAPI, Docker, RDKit, MCP |
| Agentic AutoML Platform | LangGraph ReAct agent orchestrating end-to-end tabular ML pipelines via a modular MCP tool server; five-container Docker architecture with MLflow tracking and LangSmith observability | LangGraph, MCP, FastAPI, Docker, MLflow, XGBoost, CatBoost, TensorFlow, scikit-learn, LangSmith |
| NL-to-SQL Agentic Pipeline · 🤗 demo | Hybrid BM25 + embedding retrieval with RRF fusion and execution-guided self-correction; 54.5% execution accuracy on BIRD Mini-Dev, 55% token reduction; ablation harness with failure-mode taxonomy | LangChain, Voyage AI, SQL |
| Multi-Agent Symbolic Reasoning · 🤗 demo | Solver → verifier → critic loop with LangGraph conditional edges; SymPy verifier deployed as independent MCP server over SSE | LangGraph, MCP, SymPy |
| GDL/RL Agent: Orbit Wars | GATv2 policy network with action masking, differentiable k-NN graph edges, two-stage curriculum; documented four compounding training failures including silent SB3 optimiser bug | PyTorch Geometric, PPO, SB3 |
| Domain-Adaptive LLM Trilogy | RAG → GraphRAG → QLoRA fine-tuning pipeline over my physics PhD thesis; ontology-driven knowledge graph, 2× retrieval benchmark gain, domain-adapted Qwen2.5-3B (🤗 adapter) on free-tier T4 | QLoRA, LlamaIndex, GraphRAG (NetworkX, Pydantic) |
| Deep Learning for Time Series Forecasting | LSTM vs Transformer vs AutoTFT on BTC/USDT (71k hrs); XGBoost demand forecasting on Grid India data with climate scenario simulation (+2°C/+4°C stress tests) and MLflow tracking | PyTorch, TensorFlow, NeuralForecast, XGBoost |
| Low-Resource NMT: EN→ML | MarianMT fine-tuned on BPCC + NPTEL Shiksha corpus; BLEU 15.48 → 32.61 (2× gain) on a low-resource pair where general-purpose models fail on technical vocabulary (🤗 weights, 🤗 domain-adapted weights) | MarianMT, Hugging Face |
