AI/ML Researcher & Engineer
B.E. Artificial Intelligence & Data Science · Savitribai Phule Pune University
Pune, India
I work on interpretability and evaluation of language models understanding what happens inside them, and building tools that make model behaviour measurable. Most of what I build sits between research and engineering: probing internals, designing evaluations, and turning findings into something reproducible.
I'm early in my research journey and still exploring. Interpretability is where I've spent the most time so far, but I'm actively reading and building across NLP, multimodal models, LLM systems, alignment and evaluations and my interests are still widening.
Alongside research I work on applied ML, Agentic AI and GenAI systems, and contribute to open source.
- AI Research Fellow at Single Core Labs — causal representation dynamics and RL reward-model robustness, studying latent reward hacking in open-weight models
- Open Source Contributor at Cohere Labs and GSSoC'26 — Indic-language LLM benchmarking, agentic workflows, and EU AI Act compliance tooling
Where I've worked so far
- Mechanistic interpretability — activation steering, causal patching, sparse autoencoders
- Behavioural evaluation of LLMs — sycophancy, agentic welfare trade-offs, compliance
- Applied NLP — RAG pipelines, semantic search, multilingual benchmarking
- AI governance in practice — mapping model behaviour to regulatory frameworks
What I'm exploring next
NLP research · multimodal models · LLM post-training · agentic systems · chain-of-thought faithfulness · scalable oversight
Always happy to talk about any of it.
Research & Interpretability
| Project | Description |
|---|---|
| EmotionScope | Extracts, probes, and steers emotion and sycophancy vectors across residual streams in open-weight LLMs. |
| InterpLab | Interactive workspace for logit-lens inspection, causal activation patching, and SAE probing. |
| TAC-Procure | Agentic welfare benchmark measuring implicit choice tendencies under cost pressure, mapped to EU AI Act Article 55. |
Applied AI & GenAI
| Project | Description |
|---|---|
| EKAM | An AI event operating system — agentic workflows for planning, coordinating, and running events end to end. |
| Praxis | AI-assisted coding practice platform with a teacher-in-the-loop review gate. |
| AiUniverse | Multi-agent platform that queries six language models in parallel and returns their answers side by side, surfacing where models agree and diverge. |
Machine Learning
Languages
Backend & Web
Tools & Infrastructure
- Technical Team Member, TEDxPVGCOETM
- Volunteer, National Service Scheme (NSS), PVGCOET Unit
- International Delegate, Harvard Project for Asian and International Relations (HPAIR ACONF'26)
- National Finalist, MIT-WPU Ignisia 24-Hour National AI Hackathon
- 3rd Place, HardHack Forge National Hackathon, PCCOE Pune
Currently open to internships and research opportunities
AI/ML · GenAI · NLP · Data Science · AI Research & Backend Engineering
📫 gadesaanidhi@gmail.com · LinkedIn · Discord