AD

Hi, I'm

Aritra Das

Problem solver — I bend ideas into things that work.

Move your cursor — it's a black hole.

About

Applied Scientist Intern at Microsoft and M.Tech Computer Science student at IIT Roorkee, specializing in large language models, deep learning, and applied ML research. Hands-on experience in LLM fine-tuning, agentic systems, model compression, and interpretability — delivering measurable, reproducible results. Targeting Applied Scientist, Research Scientist, and AI Engineer roles.

Experience

Applied Scientist Intern — Microsoft

May 2026 – Present · Hyderabad, India
  • Designing and building scalable evaluation frameworks to measure the quality, reliability, and safety of LLMs across diverse, real-world tasks.
  • Developing automated benchmarking and metric pipelines that surface model failure modes and behavioral regressions for data-driven decisions.
  • Collaborating with cross-functional research and engineering teams to turn evaluation insights into model robustness and trustworthiness gains at scale.

Deep Learning Research Intern — TCS Research & Innovation

May 2022 – Jul 2022 · Kolkata, India
  • Applied Lottery Ticket Hypothesis pruning to compress VGG, ResNet, and FCN models, retaining 88% accuracy via confidence-driven selective classification.
  • Built an ECG time-series classification pipeline, improving cardiac-anomaly detection by 18% over baselines.
  • Developed scalable CNN training workflows, cutting training time by 10% and improving reproducibility.

Earlier: Toppr — Mobile App Developer Intern (4+ React Native quiz modules, 100K+ users); iLRNU — Full-Stack Developer Intern (React.js, Node.js, MongoDB).

Education

M.Tech, Computer Science & Engineering — IIT Roorkee

2025 – Present · GPA 8.7/10

Coursework: Machine Learning, Deep Learning, Algorithms, Quantum Computing.

B.Engg, Computer Science & Engineering — Jadavpur University

2019 – 2023 · GPA 8.03/10

Coursework: Artificial Intelligence, DBMS, Data Structures & Algorithms.

Projects

Vision Transformers for High-Energy Physics

TensorFlow · Keras

A ViT classifying electrons vs. photons from CERN calorimeter data — 88% accuracy, 0.91 ROC-AUC, beating CNN baselines by 10–12%, with custom patch encoding and attention layers.

Agentic LLM Fine-tuning with RLHF

PyTorch · Transformers

End-to-end RLHF (reward modeling + PPO) on a 1.3B-parameter LLM. Lifted human-preference win-rate 52→68%, cut unsafe actions ~35%, and raised multi-tool task success 61→79%.

Agentic Workflows with kgraph + MCP

Python · Model Context Protocol

A suite of specialized LLM agents backed by kgraph, a persistent knowledge-graph memory exposed as an MCP server — grounding agents in cited facts and reducing hallucination across sessions.

Publications & Open Source

Skills

Languages

  • Python
  • C
  • C++
  • Java
  • SQL
  • JavaScript

ML & AI

  • PyTorch
  • TensorFlow
  • JAX
  • Hugging Face
  • Scikit-learn

LLM & Applied ML

  • Fine-tuning
  • RLHF / PPO
  • Agentic Systems
  • MCP
  • RAG
  • Interpretability
  • Model Compression

Systems & MLOps

  • Distributed Training
  • DeepSpeed
  • Docker
  • Linux
  • Git
  • Weights & Biases

Data & Cloud

  • PostgreSQL
  • MongoDB
  • FAISS
  • AWS
  • Google Cloud

Open to Collaboration

I'm open to collaborating on papers, projects, and patents. If your work overlaps any of these areas, let's talk.

LLM Evaluation & Benchmarking

Scalable eval frameworks, metric design, failure-mode and regression analysis.

Interpretability

Probing and understanding transformer internals for transparent models.

Agentic Systems & RLHF

Tool-using agents, planning loops, reward modeling and preference optimization.

Model Compression & Efficient ML

Pruning, distillation, and quantization for fast, deployable models.

ML for Particle Physics

Deep learning on CERN calorimeter data — event classification and scientific imaging.

Quantum Computing

Quantum algorithms and quantum machine learning on the physics frontier.

Contact

Open to Applied Scientist, Research Scientist, and AI Engineer roles. Reach me here: