AI research and systems, Mumbai
Language models that run in less memory, hold up under rigorous evaluation, and stay inside their limits.
Third-year B.Tech in Computer Science at SPIT Mumbai, GPA 9.4. First-author preprint on KV-cache compression. Research intern at IIIT Kancheepuram and IIT Jodhpur. Open to collaborations and internships.
Long-context inference runs out of GPU memory first. My preprint cuts the KV cache by 35.1% and keeps quality.
A result needs a test it can fail. I run full benchmarks, 16 LongBench subsets, and significance tests, not hand-picked examples.
Agent-written code runs in a sandbox with no network and hard limits, behind a server that manages its own GPU memory.
Safe control of a power grid, jailbreaks across text and video: each gets a measured answer, not a demo.
Online low-rank compression of pre-RoPE keys in grouped-query attention, with per-head adaptive rank allocation, a custom CUDA de-rotation kernel, and a randomised SVD path. Over 1,602 samples on all 16 LongBench subsets (Qwen2.5-7B-Instruct, 4-bit NF4, dual T4): 73.91% ROUGE-L win rate against the uncompressed baseline.
Python-loop SVD added 29.5% latency, so the algorithm was not the whole story: the implementation had to be batched and randomised. Head-level analysis mattered too, since a few high-entropy heads need more rank than the rest.
A differentiable residual-Dyna world model: an analytical AC power-flow base plus a learned residual, a quadratic-programming safety projection, and personalised federated learning. Early results are promising.
Safety constraints belong in the architecture, not the reward. The projection layer guarantees feasible actions, so the learned residual only has to correct the physics model.
Analysis of distilled small language models on multilingual tasks, extending Microsoft's Phi family to new evaluation settings. Equal contributor on hypotheses, experiments, and evaluation.
Probes LLMs across five attack categories and measures how jailbreak success shifts between text, image, and video delivery, compared with McNemar and Wilcoxon tests. Code is private for now.
Raw pass rates mislead. Paired tests changed which differences I trusted, and the modality gap only became visible once delivery format was held as the single changing variable.
An air-gapped AI workbench for SIH 2026. An OpenAI-compatible GPU gateway with LRU memory management and a Docker and gVisor code sandbox, plus a client with a tool-calling agent orchestrator.
Server code, client at Indra_lok on GitHubAgentic skill-bartering platform built for the JPMC Agastya Foundation. Matches learners and mentors with no manual moderation.
Live siteCivicFlow and the FAR AWAY agent find a form, infer its fields, and fill it in end to end with Playwright or Selenium.
FAR AWAY code on GitHubOffline Bluetooth-mesh disaster messaging, YOLO rockfall detection, and a PaddleOCR pipeline that recovers text from degraded scans.
All repositories on GitHub