Yash KVR

Mumbai · Coincade Studios

Yash Khanvilkar

  • Fine-tuned models.
  • Agents that retrieve.
  • Vision on device.

I fine-tune and deploy frontier and open-source LLMs, build enterprise RAG and agentic workflows with LangChain and the OpenAI Agents SDK, and run computer vision from the edge to the robot.

See the systemsDM on LinkedIn

inference

serving
  1. trainfine-tune · frontier + open
  2. servemodal · serverless gpu
  3. retrieveenterprise rag
  4. actlangchain · openai agents
  5. seedetect · track · verify

Train it. Serve it. Let an agent use it. Robotics is where the same stack grows a body.

Fine-tuningFrontier LLMsOpen-source LLMsModalServerless GPUsEnterprise RAGLangChainOpenAI AgentsComputer visionEdge inferenceFine-tuningFrontier LLMsOpen-source LLMsModalServerless GPUsEnterprise RAGLangChainOpenAI AgentsComputer visionEdge inference

01Focus

Models, vision, then agents.

Machine learning and computer vision are the core. Agentic workflows are how those models get used. Robotics is the hardest place I put them.

01 · Models in production

Machine learning

Fine-tuning and deployment for both frontier models and open-source LLMs. Inference runs on Modal, on serverless GPUs, so a model can scale without a cluster I have to babysit.

  • Fine-tuning
  • Frontier and open-source LLMs
  • Modal serverless GPUs

02 · Perception

Computer vision

Detection, tracking, and identity on the device. The same visual stack shows up in edge cameras and in the vision-language-action policies that drive a robot’s hands.

  • Edge inference
  • Detection and tracking
  • Vision-language-action

03 · Workflows

Agentic AI

Enterprise RAG apps, and multi-step agentic RAG workflows on top of them. LangChain and the OpenAI Agents SDK handle retrieval, tools, and the decision of what to do next.

  • Enterprise RAG
  • LangChain
  • OpenAI Agents SDK

Also · Robotics

A body for the model

Humanoid walking in Isaac Lab, dexterous manipulation with LeRobot and vision-language-action models, and a language-level agent that calls ROS 2. I lead the robotics software and data collection teams through live hardware demos.

02Experience

Lead developer at Coincade.

Since December 2021. The job is models, vision, and agents. Robots are where that work has to survive contact with the real world.
Illustrated portrait of Yash Khanvilkar

Yash Khanvilkar

Lead Developer — ML, Vision & Agents

Coincade Studios · Mumbai

Machine learning, vision, agents

Dec 2021 — Present

  • Fine-tune and deploy frontier and open-source LLMs on Modal, using serverless GPUs for inference.
  • Build enterprise RAG applications, then agentic RAG workflows on LangChain and the OpenAI Agents SDK.
  • Ship computer vision that detects, tracks, and verifies identity at the edge.

Robotics

  • Humanoid locomotion in NVIDIA Isaac Sim and Isaac Lab, trained to transfer onto uneven ground.
  • Imitation learning and vision-language-action policies for dexterous manipulation.
  • Lead the robotics software and data collection teams, including live hardware demonstrations.

Earlier in the same role I also shipped product software, including Jibbr, a realtime messaging app used across the company.

03Selected systems

What is actually running.

Models on Modal, retrieval workflows, and vision come first. Robotics is the same stack with motors attached.

Machine learning

LLMs in Production

Frontier models and open-source models, fine-tuned and served. Serverless GPUs so inference shows up when a request does, and goes away when it doesn’t.

In the lab

  • Fine-tuned both frontier and open-source LLMs for real workloads.
  • Deployed inference on Modal with serverless GPUs.
  • Kept the path from a trained checkpoint to a callable endpoint short.
LLMsFine-tuningModalServerless GPUsPython

Agentic AI

Agentic RAG workflows

Retrieval that doesn’t stop at one lookup. Agents plan the next step, call tools, and check what came back before they answer.

In the lab

  • Built multiple agentic RAG workflows for real tasks, not a single demo chain.
  • LangChain for the retrieval and tool plumbing.
  • OpenAI Agents SDK where the workflow needs a model that can decide and delegate.
LangChainOpenAI Agents SDKRAGLLMsPython

Enterprise RAG

Retrieval apps

RAG applications for enterprise use cases: company knowledge in, a grounded answer out, with the model kept on a leash by what was actually retrieved.

  • Designed retrieval apps around internal knowledge, not the open web.
  • Wired the model so the answer stays attached to the sources it used.
RAGLLMsPython

Computer vision

Edge vision

An automated camera system that detects, tracks, and checks identity on the device, instead of waiting on a round trip.

  • Real-time detection and tracking at the edge.
  • Identity verification in the same pipeline.
Computer visionEdge inferencePython

Robotics

Humanoids and manipulation

The same ideas on hardware. Walking policies in Isaac, imitation and vision-language-action for the hands, and an agent that calls ROS 2.

  • Bipedal locomotion in NVIDIA Isaac Sim and Isaac Lab, with domain randomization and sim-to-real transfer.
  • Imitation learning with LeRobot, π₀, and NVIDIA GR00T.
  • A language-level brain on the OpenAI Agents SDK that decomposes a task and runs ROS 2 skills.
Isaac LabROS 2LeRobotVLAReinforcement learning
Source
More on GitHub

04Toolbox

The stack I reach for.

Python for the models. LangChain and the OpenAI Agents SDK for the workflows. Modal when a GPU has to show up on demand.

Machine learning

  • Fine-tuning
  • Frontier LLMs
  • Open-source LLMs
  • Modal
  • Serverless GPUs
  • Deep learning

Agentic AI

  • LangChain
  • OpenAI Agents SDK
  • Enterprise RAG
  • Agentic RAG
  • Tool use
  • Python

Computer vision

  • Detection
  • Tracking
  • Edge inference
  • Identity
  • Vision-language-action
  • π₀

Robotics

  • ROS 2
  • NVIDIA Isaac Sim
  • Isaac Lab
  • LeRobot
  • NVIDIA GR00T
  • Sim-to-real

05Contact

A model to ship, a workflow to trust, or vision that has to be right.

I am at Coincade Studios in Mumbai. Write about LLMs, RAG, agents, computer vision, or a robot that needs the same stack.

yash.khanvilkar1@gmail.comBack to the work