Kestrel
A multi-tenant GPU compute control plane with runway-aware scheduling.
Python · FastAPI · Kubernetes · Kueue · PostgreSQL · Prometheus
LLM inference & evaluation
Hey, I’m Mohith.
I measure and build the systems that serve LLMs.
Final-year CS student and independent researcher.
Two preprints on LLM inference and eval reliability, GSoC '26 contributor at RUXAILAB , and former product developer intern at Koncepts. Currently building ServeBench, an environment for evaluating agents on LLM serving.
Experience
A few places I’ve built, learned, and taken responsibility.
- Started as an open-source contributor in Nov 2025 and was selected for Google Summer of Code 2026 (May 2026).
- Built the Focus Group Management & Analysis module end to end for RUXAILAB's open-source usability-testing platform (20 PRs, 18 merged).
- Shipped 30 PRs overall (24 merged), working with my mentor through code review.
- Added live focus-group sessions with LiveKit (SFU) and NLP-assisted thematic analysis using Python Cloud Functions on Firebase.
Vue.jsPythonFirebaseCloud FunctionsLiveKit- Shipped 165+ PRs to production on an edtech platform serving ~200K learners.
- Built the DP Mock Paper Builder end to end and the MYP eAssessment Paper Builder, generating criterion-aligned assessments with AI-generated images.
- Launched PYP: QR class-card login, a child-safety video gate, and parent-consent onboarding.
- Built a collaborative Unit Planner with edit locks and version snapshots.
- Fixed a job-queue race condition by making job claiming atomic in Postgres, and fixed LLM timeouts by streaming structured outputs.
PythonPostgreSQLNext.jsTypeScript- Completed a one-month on-site internship at NIELIT Kozhikode, focused on backend web development.
- Built web applications with Python and Django, covering models, views, templates and database integration.
PythonDjango- Founded the chapter and served as its first chair for nearly two years, building it from zero.
- Led the student team that planned and ran the chapter's technical events and workshops.
- Built a community of students around ocean engineering and technology at the university.
THE FOUNDATION
Chinmaya Vishwavidyapeeth
B.Tech, Computer Science
Recognition
- Selected as 1 of 6 presentersIndia Builds with Claude (Anthropic × Razorpay), from 2,300+ applicants
- 1st placeOpenAI Codex Community Hackathon, Calicut (for kea)
- Dean's Merit Listranked 1st in CS department
- Mintlify Open Source ProgramAntiVibe
Work
04 PROJECTSAI infrastructure, developer tools, and a few things I wished existed.
CacheKit
Finds what breaks LLM prompt caching, opens a fix PR via an agent, and checks behaviour with a quality guard before merge. On Qwen3-4B (vLLM, L4): cache-hit rate 0.3% → 80.7%, median TTFT 2.0s → 1.4s.
Python · vLLM · Modal · Next.js
Kea
Real-time incident analysis with deterministic root-cause ranking and evidence-backed AI explanations.
1st place · Codex Community Hackathon, Calicut
Python · FastAPI · Neo4j · Redpanda · Next.js · Codex
antivibe
A Claude Code skill that makes you read the code it wrote.
1,000+ GitHub stars · Part of the Mintlify OSS Program
Shell · Claude Code · MCP
Latest writing
All posts →Activity
- 0
- Contributions
- 0
- Merged PRs
- 0
- Stars earned
- 0
- Public repos
Stack
Every day
What I open without thinking about it
vLLMPyTorchHugging FacePythonTypeScriptFastAPINext.jsPostgreSQLDockerGitClaude Code
Often
Reached for whenever the problem calls for it
llama.cppModalGGUF quantizationReactKubernetesRedisTailwindSupabaseVuePrometheusVercel
When it fits
Used in anger at least once, happy to again
JavaLangChainOllamaFirebaseMySQL
Learning
- LLM serving & KV caches
- RL environments
- Post-training
- GPU kernels (Triton)
A line that stays with me