Shuvam Banerji Seal Shuvam-Banerji-Seal

About

BS-MS student at Indian Institute of Science Education and Research, Kolkata (Chemistry Major, Computer Science Minor, CGPA 8.2). My work sits at the intersection of computational chemistry, AI/ML research, and information retrieval. I co-founded two MeitY-funded deeptech startups and have published at ECIR 2026, FIRE 2025, and TREC 2024.


Current Research	Hybrid RAG architectures, DFT-based catalyst modelling, query reformulation for NLP
Learning	LAMMPS, HPC workflows, advanced DFT, Agentic AI frameworks
Looking to Collaborate	ML, Data Science, Molecular Simulations, Quantum Chemistry
Contact	sbs22ms076@iiserkol.ac.in

Philosophy

First principles, then force fields. — Every simulation is an argument about reality. I compute to test the argument, not to decorate it. If the HOMO–LUMO gap disagrees with the experiment, the gap is not wrong — the model is.

Retrieval over memorisation. — A model that cannot cite its sources is an opinion engine. Whether it is a language model or a spectroscopy reference, the answer must trace back to evidence.

Ship the negative result. — The failed catalyst, the ablation that lost, the retrieval run that underperformed BM25 — these are data. Science compounds only when the null results are public.

Research — Animated Overviews

Publications

Year	Venue	Title
2026	ECIR 2026 (34.5% acceptance)	AgriIR: A Scalable Framework for Domain-Specific Knowledge Retrieval
2025	FIRE 2025	Hierarchical Opinion Classification using Large Language Models
2024	TREC 2024 (NIST)	IISERK@ToT_2024: Query Reformulation and Layered Retrieval
WIP	To be submitted	Computational Modeling of [VO(SALIEP)(DTP)] as Water Reducing Catalyst (DFT/B3LYP)

Extended Visualizations

Attention Heads / Transformer Architecture

Molecular Orbital Spin Diagram — [VO(SALIEP)(DTP)]

Tech Skills

Molecular / HPC

Languages

AI / ML / Data

Frameworks & Tools

Interactive Lab — Synthesize the Catalyst from the DFT Study

Can you replicate the complex from the computational chemistry research? Choose wisely — wrong paths teach too.

Step 1 — Choose your central metal

You need a redox-active transition metal capable of oxo-coordination for water reduction.

Select Vanadium (V)

Correct instinct. Vanadium(IV) oxo-complexes are known water-reduction catalysts — the d¹ electron gives you EPR handle too.
- Add Salen-type ligand (SALIEP)
  
  SALIEP provides strong N/O donors that stabilise V(IV). Now choose the co-ligand.
  - Add dithiophosphate (DTP)
    
    You built [VO(SALIEP)(DTP)] — the exact target complex from the DFT/B3LYP study. Run Gaussian with B3LYP/6-311G(d,p), compute MO energies, map the water-reduction pathway. The V=O stretch should land near 965 cm⁻¹ — if it does, your geometry converged.
  -
  Add bipyridine (bpy)
  
  Interesting coordination — but bpy makes the complex too inert for proton-coupled electron transfer here. Swap in dithiophosphate (DTP).
- Add only H₂O as ligand
  
  Water alone will not chelate stably — it exchanges too fast. Try a stronger donor like SALIEP.

Select Iron (Fe)

Iron catalyses plenty of chemistry — but this specific DFT study uses Vanadium. Try again.

Select Zinc (Zn)

d¹⁰ — diamagnetic, redox-silent. No unpaired electron, no EPR signal, no water reduction here. Pick a metal with available d-electrons.

Interactive Lab — Build the Hybrid RAG Pipeline (ECIR 2026)

Design the retrieval architecture step by step — same decisions I made for the IISER-K intranet system.

Step 1 — Choose your retriever

Dense retrieval only (vector embeddings)

Handles semantic similarity well, but misses exact keyword matches for rare technical terms. Upgrade?
- Add BM25 sparse retrieval in parallel
  
  Now you have hybrid retrieval. How do you merge the two ranked lists?
  - Reciprocal Rank Fusion (RRF)
    
    Perfect. RRF combines ranked lists without score calibration. Add HyDE (hypothetical document embeddings) for query refinement and wrap in a Streamlit UI — you have replicated the ECIR 2026 / IISER-K intranet system.
  -
  Weighted linear score combination
  
  Works, but requires per-domain score calibration — brittle across corpora. RRF is more robust. Try again.

BM25 sparse only

Fast and interpretable, but misses paraphrases. Upgrade to hybrid for better recall.

Fine-tune a cross-encoder re-ranker on 50 labelled queries

Legitimate architecture — but with 50 labels you will overfit in an afternoon. Hybrid + RRF gets you further with zero labels. Try the label-free path first.

Interactive Lab — Run the MD Simulation (LAMMPS)

Solvate the vanadium complex and keep it intact for 500 ns. Three choices, one stable trajectory.

Step 1 — Pick the ensemble

NVT (fixed volume, thermostat only)

Stable and simple — but with fixed volume you cannot model the solvent density change as the complex relaxes. Fine for a first equilibration, though. Proceed.
- Thermostat: Nosé–Hoover
  
  Smooth, deterministic, canonical ensemble. Now the integration timestep:
  - 2 fs with constrained bonds (SHAKE)
    
    Stable trajectory. Constraints remove the fast X–H vibrations, 2 fs resolves the V=O stretch, and over 500 ns the RMSD plateaus near 1.2 Å — complex intact, no dissociation. This is the production setup.
  -
  5 fs bare (no constraints)
  
  Blown up. The O–H stretch oscillates at ~10 fs period — a 5 fs step integrates it as energy gain and your simulation detonates within picoseconds. Constrain bonds and drop to 2 fs.
- Thermostat: direct velocity rescaling every step
  
  It will hold the temperature — but it distorts dynamics and gives wrong fluctuation statistics. Nosé–Hoover samples the canonical ensemble properly.

NVE (no thermostat)

Pure Newtonian dynamics — beautiful, but any drift from the equilibrated state never gets corrected. Use NVT for production.

GitHub Stats

Every stat service below was verified alive (HTTP 200, valid SVG) at embed time. Dead endpoints — github-readme-stats (deployment paused), profile-trophy and contributor-stats (disabled), streak-stats herokuapp (shut down) — were removed in the v2 redesign rather than left rendering error images.

Research Workflow

flowchart LR
    A[Problem Statement] --> B[Literature Review]
    B --> C[Dataset & Baseline]
    C --> D[Model Design]
    D --> E[Experiments & Ablations]
    E --> F{Results Satisfactory?}
    F -- No  --> D
    F -- Yes --> G[Paper Writing]
    G --> H[Peer Review]
    H --> I[Publication ECIR / FIRE / TREC]
    I --> J[Open-Source Release]

    style A fill:#1e3a5f,stroke:#58a6ff,color:#c9d1d9
    style I fill:#2d1f4a,stroke:#d2a8ff,color:#c9d1d9
    style J fill:#1f3a2d,stroke:#56d364,color:#c9d1d9

Navigation Menu