Inbasekaran S.

RESEARCHER · READER · AUTHOR

Inbasekaran S.

LLM inference.
Memory systems.

I’m a third-year undergraduate at SRMIST, working on LLM inference and memory systems.

I study context pruning and KV-cache management, with a growing interest in information theory.

Read my research Get in touch
Hand-drawn portrait of Inbasekaran S
Based in India AI systems & LLM inference
What I’m working on Founder, ILRL Labs ↗

Research

My research.

I’m interested in how language models use time and memory. My work focuses on context pruning, KV-cache management, and the information theory behind those choices. I also work on zero-trust software verification.

01 CONTEXT PRUNING

EntroPrefill

Rényi-Guided Context Pruning with Conditional Stability Guarantees for Retrieval-Augmented Generation

Inbasekaran S
arXiv:2610.09757v1 [cs.LG] · Oct 2026 -43.1% TTFT Read paper ↗ Google Scholar ↗

I study which parts of a long context can be removed while an LLM is still processing the prompt. EntroPrefill uses Rényi entropy to propose what to prune, with checks on the attention mass that is removed.

Technical notes
  • I implemented a training-free context-pruning method that reduced time to first token (TTFT) by 43.1% in the reported experiments.
  • The scoring uses Order-2 Rényi entropy on GQA-pooled attention, alongside conditional transformer perturbation bounds.

02 KV-CACHE MANAGEMENT

Page-EntroKV

Hardware-Aligned, Entropy-Weighted KV-Cache Eviction under Grouped-Query Attention

Inbasekaran S
arXiv:2610.03135v1 · Oct 2026 UOR 1.000 · 100% Needle Recall Read paper ↗ Abstract ↗

I work on how an LLM should use a limited KV cache. Page-EntroKV scores and evicts memory at the page level, matching the layer, group, and page units used by serving systems.

Technical notes
  • The framework keeps the Union Overhead Ratio (UOR) at 1.000. Head-independent replay used up to 4.75×–5.1× as much cache on Qwen2.5-1.5B-Instruct.
  • It removes a 13× attention-sink masquerade and reaches 100% needle recall at a 20% KV budget in the reported test; arithmetic-mean pooling reaches 0%.
  • Page allocation uses Rényi-2 collision entropy and Hamilton–Hare largest-remainder rounding.

03 SOFTWARE VERIFICATION

VendorChain

Zero-Trust B2B Supply-Chain Verification Engine

Inbasekaran S · Undergraduate Research Opportunities Program (UROP) at SRMIST
UROP Research Zero-Trust & Cryptography

I built VendorChain to check software as it moves through a supply chain. It combines zero-trust file intake, signed software bills of materials, and a private blockchain record of verification decisions.

Technical notes
  • I built a NIST SP 800-207-aligned intake pipeline with binary magic-byte validation, AES-256-GCM envelope encryption, and in-memory zeroing.
  • I used Syft to extract CycloneDX software bills of materials and Cosign to sign them, so an artifact’s origin and integrity can be checked.
  • I deployed a private Hyperledger Fabric network with Go chaincode to record SHA-256 artifact hashes, policy decisions, and timestamps.

A little about me

A little
about me.

Portrait of Inbasekaran S

I’m studying Computer Science and Engineering (Cyber Security) at SRM Institute of Science and Technology. I’m in my third year, with a CGPA of 8.66/10, and expect to graduate in May 2028.

My research centres on LLM inference and memory systems, particularly context pruning and KV-cache management under grouped-query attention. I also build AI agent workspaces and zero-trust software verification systems.

Outside that work, I write. I’m the author of Neither Living Nor Dying: Reflection of the Middle.

Outside research

I’ve always loved reading.

I often read about human psychology, psychoanalysis, and finance.

Lately, I’ve been spending more time on mathematics, particularly information theory, and thinking about inference for AGI.

Recognition

Awards & recognition.

Hackathon Winner
Winner, Raccoon AI Hackathon
I won first place after building and demonstrating an applied AI system.
National Finalist
Shark Tank Nationals
I was selected as a national finalist and presented an AI product prototype to an investor panel.
Top 10 Finalist
Hack Rush Hackathon
I placed in the Top 10 after building a working software MVP in a 24-hour sprint.
Top 0.68% India
Naukri Campus Young Turks
I scored in the 99.32 national percentile in Naukri Campus Young Turks’ coding and technology assessments across India.
2026
Oracle Cloud & AI 4x Foundations Associate
I earned four Oracle Cloud & AI Foundations Associate certifications in 2026.
2025
Prompt Engineering for Generative AI · NPTEL Silver Rank · Lean Six Sigma Green Belt
In 2025, I completed Prompt Engineering for Generative AI through DeepLearning.AI / Coursera, earned an NPTEL Silver Rank in Fundamentals of C/C++/Java/DBMS, and earned the IASSC Lean Six Sigma Green Belt.

Experience & education

Where I’ve learned.

Jul – Aug 2026
AICTE AWS Gen AI Intern Virtual
I worked through generative AI labs and planned model training and evaluation exercises using Amazon SageMaker and AWS Academy, documenting accuracy and performance metrics.
Apr – Jun 2026
AICTE Google Virtual Intern — AI/ML Virtual
I completed hands-on machine learning exercises, practical workflows, and applied AI/ML projects.
2024 – 2028
SRM Institute of Science and Technology (SRMIST)
B.Tech in Computer Science & Engineering (Cyber Security) · CGPA: 8.66 / 10 (Expected May 2028)
Relevant Coursework: Data Structures & Algorithms, Database Management Systems, Computer Networks, Operating Systems, Artificial Intelligence, Information Security, System Design.

People & teamwork

Working together.

Aug 2025 – Aug 2026
Creatives & Sponsorship Lead, Founders Club SRMIST
I led technical, creative, and outreach work for 5+ university startup events. Improving how the teams communicated helped reduce deliverable turnaround time by 25%.
Aug 2025 – Aug 2026
Technical Member, E-Cell SRMIST
I built websites and promotional materials for campus entrepreneurship events.
Aug 2024 – Present
Corporate Member, IEEE Student Branch SRMIST
I worked on corporate outreach and sponsorships to bring industry partners into technical events.

Beyond research

Writing.

Lately

Recent updates.

Research & updates

Get in touch

Let’s talk.

If you’d like to talk about my research, work on something together, or recommend a book, I’d love to hear from you.

Send me an email inba.research.ilrl.labs@gmail.com