Final year CSE (AI/ML) at IIIT Nagpur. I build AI systems that hold up outside the notebook: real constraints, real data, real failure modes.
Right now I'm deep into agentic systems, LLM post-training, and multi-agent RL.
Agents & RL:
- Multi-agent dispatch system for a 20-node city POMDP. GRPO on Qwen3-4B, adversarial curriculum, radio delays, ghost calls (DispatchR)
- Multi-agent research assistant. LangGraph pipelines over arXiv + Semantic Scholar (ArXiv Scout)
- Prompt versioning + eval platform with A/B testing and LLM-as-judge scoring (PromptLab)
Research:
- XAI framework for diffusion-based medical classifiers. Interpretability across 1000 denoising timesteps (repo)
- DiffMIC-v2 from scratch. Diabetic retinopathy grading with dual-granularity conditioning, 84.1% on APTOS-2019 (repo)
- Teaching Mistral-7B to reason with GRPO. ZS 41% → SFT 45% → GRPO 52% on GSM8K (experiment)
Systems:
- RAG-based CS interview prep platform. Structured output pipeline, offline eval cut malformed JSON from ~12% to <1% (Recall.cs)
- Real-time PPE compliance monitoring. Event-driven, EMA fusion, stable under occlusion (SentinelVision)
- NL→SQL with chain-of-thought reasoning, 78% execution accuracy on Spider (NL2SQL)
- Medical report parser with dual OCR + Gemini pipeline (BloodParser)
For myself:
- Knowledge curation tool with contradiction detection. Because saving tabs isn't the same as learning (Noesis)
Languages
ML/AI
Frameworks & Tools
- Interpretability, not just accuracy
- Systems that handle real constraints: occlusion, latency, limited compute
- Honest reporting, including the parts that didn't work
- READMEs that actually explain the tradeoffs
tejas@tejasgarg.in • Portfolio • LinkedIn
📍 Nagpur • IIIT Nagpur '27