AI for biotech. Graphs, geometry, and single-cell data.

AI Genomics Research Intern 1 month

Erdős AI Lab

• Built EG-IPG, an evidence-gated GNN over an 8M-edge protein-interaction graph, with gating applied at the level of individual edges.
• Combined ESM-2 and PINNACLE embeddings to predict CRISPRi Perturb-seq response in CD4⁺ T cells across roughly 11.5K targets.
• Wrote the training and evaluation pipeline: multi-GPU sweeps, a test suite, and a run registry.
• Ran preregistered multi-seed ablations on leakage-safe target-OOD splits with family-wise error control. The result was a bounded null, written up for AAAI 2027.

• 8Mgraph edges
• 11.5KCRISPRi targets

Adversarial auditing of K-NAF privacy

IIT Madras · Privacy in AI course project

• Built an auditing pipeline that attacks K-NAF decoding guarantees with targeted prompts, across optimization, validation, stress tests, and held-out evaluation, over several global K budgets and model variants.
• Measured certified privacy risk with U_EBB, effective budget, and ρ. Both settings stayed certified in aggregate.
• At k = 3 the audit transferred better to held-out data, separated attacks more sharply, and stayed more reliable than at k = 5.

Research Intern paper under review

Kansas State University · Prof. Bala Natarajan

• Built PAIGT, a position-aware inductive graph transformer for scRNA-seq: GraphSAGE with degree, shortest-path, and path-attention encoders.
• Added six biological-prior modules with 18 CLI controls, injectable at four points in the network.
• Kept the pipeline leak-free by splitting before HVG selection and fitting HVG, kNN, and prior standardization on training data only, with scTEP ensemble inference.
• Swept 12,576 configs across 5 seeds, 23 arms and 7 metrics, on 4 GPUs over 18 cohorts.

• 88,032runs swept
• +4.3 ppaccuracy
• +0.59 / +0.48ARI / NMI

AI Research and Development Intern

Photios

• Fine-tuned speech recognition models for transcribing retail sales conversations, after a literature review and a comparison of ASR architectures.
• Built a data cleaning pipeline that uses FastText to strip Sanskrit words out of Hindi datasets.
• Fine-tuned an N-Gram language model on top of the AI4Bharat Conformer, which brought word error rate down.

Student Researcher

SENAI, IIT Madras · Prof. Raghunathan Rengaswamy

• Worked on hyperbolic Grad-CAM and activation sparsity in hyperbolic and hybrid CNNs, on the question of how to interpret a model whose representations do not live in flat space.
• Those representations exposed hierarchical, part-whole structure in images.
• Gave talks on geometric deep learning, disentanglement via topology, the topology of deep networks, and PCA in ML research.

Student Researcher

RBCDSAI, IIT Madras · Prof. Karthik Raman

• Modelled reaction-protein relationships with a seq2seq pipeline: ESM-encoded proteins feeding a BART-style cross-attention decoder that generates every SMILES sequence for a given reaction.
• Tuned tokenization and data partitioning to raise the share of chemically valid SMILES.

arXiv preprint · 2605.28001

An Empirical Audit of k-NAF Budget Accounting for Anchored Decoding

How reliable the k-NAF privacy certification is once you attack it with adversarial prompts, and how the budget accounts for anchored decoding.

EMNLP 2025 · FinNLP Workshop

Enhancing Financial RAG with Agentic AI and Multi-HyDE

A dynamic agentic retrieval pipeline for SEC-10K filings: hybrid retrieval, safety guardrails, and source attribution a reader can check. Out of Inter-IIT Tech Meet 13.0.

ICCV 2025 · Beyond Euclidean Workshop · pp. 706-714

Sparse Hyperbolic Convolutional Networks with Enhanced Object Localization via GradCAM Analysis

GradCAM and activation sparsity applied to hyperbolic CNNs, which surfaces the part-whole hierarchy the geometry was chosen to hold. From SENAI, IIT Madras.

GeoNLI Top 8 of 23

Inter-IIT Tech Meet 14.0 · ISRO

An offline, on-prem vision language model for remote sensing: captioning, semantic and binary and numeric VQA, and visual grounding. I led six people and we had fifteen days.

• Curated multi-resolution datasets from VRSBench, RSVLM-QA and XLRS, then synthesized adversarial and counting or area VQA on top.
• Fine-tuned Ovis2.5-9B with LoRA for captioning and VQA.
• Built a hybrid grounding stack: Ovis intent routing into SAM-3 segmentation, OpenCV mask metadata, then oriented boxes and tool-based geometry. VQA runs the same way.

Repository

Dynamic agentic RAG for SEC-10K filings Top 6 of 23

Inter-IIT Tech Meet 13.0 · Pathway

A retrieval pipeline for financial documents that adapts to the query, refuses what it should, and shows its sources. Built on Pathway's vector store.

• Engineered Multi-HyDE, a hybrid retrieval mechanism over multi-query and hypothetical document embeddings.
• Added PDF parsing, Llama Guard safety guardrails, and source attribution.
• Deployed on Azure with Docker and FastAPI, then evaluated against GPT-4o and Gemini-1.5 on semantic similarity, recall, factual correctness, faithfulness and ROUGE.

Paper Pathway