Skip to content
PUNE--:--
IndexOpen to work

Kedar
Sathe.

@wtfkedar
Role
AI Infra · ML Systems · SDE
Based
Pune, IN
OSS
mermaid-js · vLLM
Focus
Systems · Serving · Infra

Most of my work lives in the gap between a trained model and a system someone can actually depend on — the service layer, the data plumbing, the auth and access control, the part nobody demos but everyone calls you about at 2 AM.

I help maintain mermaid-js and have patches in vLLM, Weaviate, and DeepMind's AndroidEnv. Open source is where I learn best — reading other people's production code teaches more than any course.

Right now I'm building microservices in Go, wiring up RAG pipelines that know when they don't have an answer, and occasionally writing RLHF training loops from scratch because I wanted to understand what the frameworks actually do.

Book 30 minuteskedarsathe49@hmail.comRésumé ↗
fig.1momentum SGD across a non-convex surface · three minima, one of them worth finding

Trajectory

3 positions

Building the infrastructure, systems, and platforms behind AI.

  1. Jun 2026 → now

    Remote · Mumbai

    ML / Software Engineer

    Autonex AI 360● active

    • Architected and built DealBridge from the ground up in Go — a microservices platform unifying the complete GTM-to-engineering lifecycle.
    • Owned system end-to-end: service architecture, authentication, RBAC, authorization, database architecture, schema design, APIs, business workflows, and scalability.
    • Designed workflows spanning sales, account management, leadership, engineering management, and engineering teams into a single operational pipeline.
    • Made architectural decisions around service boundaries, data ownership, access control, query patterns, and horizontal scalability.
    • Reworked PM Portal to handle 300–400+ daily concurrent users, optimising large request/response payloads, growing datasets, and production traffic.
    • Contributed to system design, backend architecture, database/query optimisation, and scalability improvements across the platform.
    • Go
    • Microservices
    • RBAC
    • PostgreSQL
    • Docker
    • System Design

  2. Jan 2026 → May 2026

    Pune

    Software / ML Engineering Intern

    Unique Schools App

    • Worked on migrating a production education platform from native Java + PHP to React Native + Node.js, modernising both mobile client and backend architecture.
    • Designed the AI workflow infrastructure powering AI interactions across mobile and web — user/request IDs, query routing, task scheduling, workflow state, and resolver assignment.
    • Built a NotebookLM-style internal RAG platform over vector databases and page-level document indexing for engineering and management knowledge retrieval.
    • Connected document ingestion, indexing, retrieval, and LLM reasoning into a single workflow over client documents, BRDs, POCs, and technical documentation.
    • Focused on retrieval quality, contextual grounding, and document-level traceability rather than simply wrapping an LLM around a search box.
    • React Native
    • Node.js
    • LangGraph
    • RAG
    • pgvector
    • Python

  3. 2024 → now

    Remote

    Engineer · Maintainer · Contributor

    Open Source, ML Systems & Freelance● active

    • Collaborator and active maintainer on Mermaid.js, a 90K+ star open-source project — review/triage PRs, investigate bugs, and coordinate issues across contributors.
    • Contributed upstream to vLLM (LLM inference/serving), Weaviate (vector/object database), and Google DeepMind AndroidEnv (RL research infra) — all merged.
    • Built end-to-end classification and regression ML systems and pipelines using MLflow and MLOps practices across the full lifecycle: data → training → evaluation → deployment → inference.
    • Built 3 end-to-end full-stack e-commerce platforms for clients with AI chatbot integrations, owning frontend, backend, APIs, databases, auth, and deployment.
    • TypeScript
    • PyTorch
    • MLflow
    • vLLM
    • Next.js
    • Docker
    • Python

Proof of work

9 shipped

Things I built because the question wouldn’t leave me alone.

Small-LLM-RLHF screenshotFeatured

Distributed training, from scratch

Small-LLM-RLHF

3D-parallel RLHF training for small language models, written from the ground up on PyTorch and Triton — no training framework underneath it.

Hand-rolled tensor, pipeline (1F1B), and data parallelism with ZeRO-1; custom Triton FlashAttention-2 kernels; a PagedAttention inference engine with its own block manager; and an async PPO loop that overlaps rollout generation on one GPU set with backprop on another.

  • PyTorch
  • Triton
  • CUDA
  • RLHF
  • 3D Parallelism
2026Source ↗
GroundedRAG screenshotFeatured

Retrieval that admits it doesn't know

GroundedRAG

Local multi-agent RAG on LangGraph, Ollama, and pgvector. Hybrid search with RRF fusion, and a refusal path for when retrieval is too weak to answer.

  • LangGraph
  • Ollama
  • pgvector
  • Python
2026Source ↗Live ↗
Ring 0Featured

Ring 0

Cybur

Linux kernel module that intercepts ARP traffic and cross-references a DHCP snooping table to stop poisoning attacks.

  • C
  • LKM
  • Kernel
  • Security
2024Live ↗

More things

  1. 2025Snake's & The Golden AppleDeep Q-learning, competitivelyMulti-agent snake trained with Double and Dueling DQN over prioritised experience replay — 25-dimensional state, 50K-step buffer.
  2. 2025LazyDocs1.2k+ npm installsCLI that reads a project's structure and writes its documentation, built for people who live in the terminal.
  3. 2025PitchForgeOutreach, tailoredGenerates cold emails, cover letters, and resume edits from a parsed PDF — coordinate-based extraction rather than flattened text.
  4. 2025Pelvix AIMultimodal, session-onlyChat and image generation over DeepSeek-R1 and Flux. Nothing is persisted, nothing is tracked — conversations die with the tab.
  5. 2025X-BirdBrowser extensionContextual reply generation on X, with tone and intensity controls running against Qwen2.5-72B.
  6. 2024FreeLixWhisper-large-v3, freeTranscription and multilingual translation with no paywall and no account — open source end to end.

More at github.com/kedar49

Instruments

36 tools
Modelling
  • PyTorch
  • TensorFlow
  • Triton
  • Hugging Face
  • scikit-learn
  • PEFT · LoRA · QLoRA
  • Unsloth
  • RLHF · PPO
  • Deep Q-Learning
Serving & infra
  • vLLM
  • Triton Inference Server
  • Docker
  • Kubernetes
  • GCP
  • Linux
  • GitHub Actions
Retrieval & data
  • LangChain
  • LangGraph
  • LlamaIndex
  • pgvector
  • PostgreSQL
  • MongoDB
  • Apache Spark
Languages
  • Python
  • TypeScript
  • JavaScript
  • SQL
  • Bash
  • Go
Interfaces
  • Next.js
  • React
  • React Native
  • FastAPI
  • Node.js
  • Tailwind CSS
  • Astro JS
fig.2 row-wise softmax over drifting logits — the weights always sum to one, which is the only promise attention makes.

Record

papers · contests · commits

Some of it is peer reviewed. Most of it is just committed.

  1. 2024AIR 1 · eYantra Stage 1, IIT BombayAwardTop 10 worldwide in Stage 2.
  2. 2024Enhanced Railway Safety SystemPaperPublished, IEEE.
  3. 2023Alzheimer's diagnosis from structural MRIPaperSVM-driven classification. Published, IJRASET.
  4. 2022Social Distancing ID CardPaperPublished, Scopus-indexed journal.
  5. 2022–26B.Tech CSE (AI/ML) · VIT PuneStudyCGPA 8.03. Deep learning, computer vision, NLP, distributed systems.
  6. —Fundamentals of Deep LearningCertNVIDIA. Plus NLP & Deep Learning, DeepLearning.AI.
Commit density@kedar49
fig.3 one cell per day over the last year, on GitHub’s own green scale.

Writing

long form

I write things down so I have to actually understand them.