class AritraDas(AppliedScientist):
def __init__(self):
self.role = "Applied Scientist Intern @ Microsoft"
self.studying = "M.Tech CSE, IIT Roorkee"
self.mission = "Measure what models really do — then make them better."
self.focus = ["LLM evaluation", "agentic systems",
"interpretability", "model compression"]
self.physics = ["ML for particle physics (CERN)", "quantum computing"]
def currently_building(self) -> str:
return "Scalable eval frameworks that surface LLM failure modes at scale."- LLM evaluation at scale — automated benchmarking & metric pipelines that catch failure modes and regressions before they ship.
- Agentic systems & RLHF — tool-using agents with planning, self-correction, and preference optimization (lifted multi-tool task success 61 → 79%).
- Interpretability — probing transformer internals for transparent, trustworthy models.
- Physics frontier — Vision Transformers on CERN calorimeter data (0.91 ROC-AUC) and quantum machine learning.
| Project | What it does |
|---|---|
| Agentic LLM + RLHF | End-to-end RLHF (reward model + PPO) aligning a 1.3B agent — win-rate 52→68%. |
| QViT | Vision Transformers → Quantum Vision Transformers for image classification. |
| Quantum Particle Transformer | Quantum transformer for high-energy physics analysis at the LHC. |
| QRL | Quantum representation learning via contrastive loss on a variational circuit. |
LLM evals · agentic systems · interpretability · model compression · ML for particle physics · quantum computing — papers, projects, or patents.