Skip to content

Experience

Research work with real systems, real experiments, and publication-aware boundaries.

These two entries are the CV experience anchors: Materia as the formal research internship and DKL-BO as the supervised undergraduate research project. Both are paper-bound, so the website uses case studies, validated results, and safe showcase assets instead of public code or live demos for now.

Experience #1Embargoed - first-author manuscript in preparation

Research Intern

Materia - Talk to a Materials Simulation Engine

Built an agentic materials-simulation platform that turns natural-language requests into real scientific workflows: VASP/DFT input generation, structure utilities, ML-potential relaxation, MD, elastic, phonon, and NEB flows across 23 tools.

Organization
CMC Lab, IIT Guwahati
Supervisor
Dr. Kalishankar Bhattacharyya
Duration
May 2026 - Jul 2026
Mode
IIT Guwahati - On-site
Ownership
Sole engineer of the full ~26K-LOC platform and first author on the accompanying manuscript, working with a PhD collaborator under Dr. Kalishankar Bhattacharyya.

Selected Work

  • Built a full-stack AI simulation platform (~26K LOC) with FastAPI, React 19, Electron, Celery/Redis, JWT auth, Fernet-encrypted API-key storage, Docker, and GitHub Actions CI.
  • Designed a 23-tool native function-calling agent on a Gemini 2.5 Flash -> Groq Llama-3.3-70B -> local Ollama failover chain, validated at 100% tool-selection and 100% argument accuracy (2.09s mean latency) over a curated 39-prompt suite.
  • Benchmarked 6 pretrained ML interatomic potentials (MACE, MatterSim) against Materials Project DFT across 60 relaxations, reaching 2.65% mean volume error and 8.3% bulk-modulus error.
  • Split deployment into a torch-free cloud web edition (15 instant tools) and an Electron desktop build (heavy GPU sims) behind a single feature gate.
AI EngineeringFull-Stack SystemsML for ScienceResearch Software
Open case study

Embargo-Safe Showcase

  • 60-90 second private screen recording: chat request, tool run, structure viewer, downloadable VASP files.
  • Validation plots: volume parity, bulk-modulus parity, scaling, and throughput.
  • Architecture diagram: agent, tool registry, instant tools, Celery jobs, provider failover, deployment surfaces.

Honesty Guards

  • No public GitHub or live demo until the paper is published.
  • Use benchmark wording for validation results, not production telemetry.
  • Desktop edition is described as in progress, not shipped.
Experience #2Embargoed - first-author manuscript in preparation

Research Intern

DKL-BO - Smarter Materials Discovery with Deep Kernel Bayesian Optimization

Designed and evaluated a CGCNN plus Gaussian-Process active-learning system for pool-based 2D-materials discovery, locating the best candidate in a 3,351-material search space while evaluating only ~3% of the pool.

Organization
CMC Lab, IIT Guwahati
Supervisor
Dr. Kalishankar Bhattacharyya
Duration
May 2026 - Jul 2026
Mode
IIT Guwahati - On-site
Ownership
Designed and implemented the full pipeline and experiments from a professor-advised research concept; first author on the manuscript in preparation.

Selected Work

  • Designed a Deep Kernel Learning surrogate with a CGCNN graph encoder, Gaussian Process uncertainty, and UCB / Expected-Improvement acquisition.
  • Located the global optimum within ~3% of evaluations - 6.3x more efficient than random search across the fixed candidate pool.
  • Validated reproducibility across 30 seeds and a second 10K-crystal SNUMAT dataset with paired Wilcoxon tests (p approx 0.002), plus uncertainty calibration (Coverage@95 = 0.93) and beta-sweep analysis.
  • Built a ~9.2K LOC Hydra-configurable research pipeline with LMDB graph caching, pool-embedding reuse, calibration metrics, and 5 focused test suites.
Data ScienceBayesian OptimizationGraph MLResearch Engineering
Open case study

Embargo-Safe Showcase

  • Static result plots: best gap over cycles, cumulative top-10% discovery, and summary dashboard.
  • Written case study: problem, DKL/BO method, results table, implementation role, and reproducibility checks.
  • Deferred post-publication demo: Hugging Face Space with a CSV-driven convergence animation.

Honesty Guards

  • Do not publish GitHub or a live demo until the paper is published.
  • Describe this as pool-based active learning over a fixed database, not wet-lab discovery.
  • Credit the advisor for concept and supervision; claim implementation and experimentation.