Applied ML engineer. I build LLM agents and put controls around them so they fail safely.
Open to ML Engineer, AI Engineer, and Data Scientist roles in the US. 📫 rohanjain2312@gmail.com
Agent reliability, mostly. Retrieval that returns the right thing, recovery loops that do more than retry, gates with real thresholds, and evals that catch a regression before a user does. I care about the engineering around the model more than the model itself.
Before this I spent three and a half years at Goldman Sachs building controls for humans — trade-risk classification, OCR/KYC pipelines, KPI analytics that replaced three FTEs of manual reporting. Same problem, different operator.
| Project | What it does | Result |
|---|---|---|
| Loan Servicing Agent | Multi-agent system for syndicated loan documents, ingests a Credit Agreement or Notice PDF and executes the lifecycle action | Confidence-gated human review, deterministic ACT/360 validation, append-only audit log |
| Self-Healing Code Agent | Generates Python, adversarially tests it, diagnoses failures via a ReAct debugger, repairs iteratively | 37% → 87% across 8 benchmark tasks, same model throughout |
| GraphBench | Benchmarks GraphRAG against GNN-RAG on multi-hop QA with the scaffold held fixed | pip install graphbench-kg |
| ToolSmith | Post-training Qwen3-4B for tool-calling via LoRA SFT and step-level GRPO with verifiable sandbox rewards | In progress |
| FinCompress | Compression study on FinBERT: structured pruning, distillation, INT8 quantization | 9.1× smaller, Macro F1 improved after pruning 50% of attention heads |
- Harness Engineering: The Part of the Agent That's Actually Yours — five places my own agent harness was the problem, with numbers from the repos above
- I Pruned Half of FinBERT's Attention Heads — and It Got Better — a compression result I had to re-run before I believed it
MS in Applied Machine Learning, University of Maryland. Previously Goldman Sachs, where I went from STEM intern to Business Intelligence Associate over three and a half years. IEEE-published on Alzheimer's MRI classification.
Stack: Python · PyTorch · LangGraph · Postgres/pgvector · Neo4j · Docker · AWS · GitHub Actions
Email · LinkedIn · Medium · Hugging Face