Skip to content

Research & Engineering

Sustainable AI-assisted engineering, from prototype to production.

What I'm building

AI-assisted engineering needs systems that keep work reviewable, context disciplined, and humans in control. I build and write at this intersection: context quality, feedback loops, backpressure, agent security, observability, and production guardrails.

Published research

Peer-reviewed conference paper · Springer, Cham ·

Understanding GPU-Level Bottlenecks in Large Language Model Inference

Shubhanshu Kushwaha · Mamata Samal · Siddhant Khare

ICDEC 2025 proceedings · Lecture Notes in Networks and Systems, vol. 2003 · pp. 211-224

My first public research paper, based on LLMTraceFX, the open-source profiler I built. It examines memory bandwidth, GPU interconnects, kernel overhead, and prefill/decoding.

My contribution
Software · Conceptualization · Investigation · Methodology · Writing - original draft · Validation
Read paper View LLMTraceFX code

Currently exploring

  • Review burden - keeping generated work inspectable before it reaches humans
  • Context discipline - deterministic input quality for LLM workflows
  • Backpressure - fast feedback loops, typed boundaries, tests, and human approval where it matters
  • Agent observability - runtime signals and audit trails that keep agent workflows inspectable
  • Governable workflows - policy boundaries, tests, and human approval where they matter
  • Writing - turning AI fatigue into practical systems guidance

Focus areas

1

Context Discipline

"How do teams keep AI systems useful when context grows, duplicates, and drifts?"

Clean and deduplicate context before it reaches the model. Use deterministic systems where possible so humans can inspect what changed.

DistillContextLabTokenVMThe Agentic Engineering Guide
2

Review Burden & Backpressure

"How do teams stop AI-generated work from overwhelming the review queue?"

Fast feedback loops, type systems, tests, and clear boundaries should catch low-signal mistakes before they reach a human.

The Agentic Engineering GuideThinkBudgetLLMTraceFXGitHub1s
3

Production Guardrails

"What systems keep AI-assisted software work governable in production?"

The goal is not more autonomous output. The goal is work that stays inspectable, bounded, and compatible with how real engineering teams operate.

OpenFGADistillContextLabGitHub1s

Products & Prototypes

Context Engineering

Distill

Deterministic context deduplication for LLMs in ~12ms

Writing

The Agentic Engineering Guide

33 chapters on context, backpressure, workflows, and building with AI

Open Source

OpenFGA

Google Zanzibar-style authorization (Core Maintainer)

Developer Tools

GitHub1s

One-second code reading for GitHub repos (Maintainer)

Context Engineering

ContextLab

LLM context engineering toolkit - analyze, compress, visualize

Memory Systems

TokenVM

KV-cache virtual memory with page-based eviction

Backpressure

ThinkBudget

Adaptive inference budget controller for reasoning workflows

Performance

LLMTraceFX

GPU-level LLM inference profiler with AI-powered insights

Working with me

Systems engineering experience meets research curiosity.

  • Ships: research prototype to production. Built Distill, maintain OpenFGA
  • Systems depth: production infrastructure at Ona (formerly Gitpod)
  • Open source: OpenFGA maintainer (CNCF), GitHub1s maintainer
  • Cross-domain: security, distributed systems, ML infrastructure
  • 80+ articles on siddhantkhare.com/writing

Let's talk

Book a call Resume Research statement