Yash Rane — AI engineer · business ops · Mumbai

I build AI agents that actually ship.

Fourth-year B.Tech AI & Data Science student in Mumbai. I build RAG pipelines, multi-agent systems and the full-stack products around them, from first principles, and put them in front of real users on AWS. Right now: agents that have warmed 250+ real-estate leads.

See the work Read the notes Résumé ↗

Currently

AI Engineer & Business Operations, Internovo Ventures

Studying

B.Tech AI & Data Science, VCET · class of 2027

Motto

Life rewards action.

numbers I can defend

cold → warm, WhatsApp agent

leads warmed

agentic RAG on Bedrock

retrieval accuracy

LangGraph query rewriting

−0%

hallucination rate

freelance backends on AWS

concurrent requests

multi-PDF, end to end

<0s

RAG latency

since July 2024

public repos

selected work

Systems that left the notebook.

Agents, retrieval and the product around them. Each one was used by someone other than me.

01Multi-agent research orchestration

ResearchOps-AI

A research team made of agents, with a judge in the room.

Work is split across specialised agents, run in parallel, and graded before any of it is trusted.

Planner, researcher and critic agents split a question into sub-tasks and run them in parallel.
An LLM-as-a-Judge scores every output against a rubric before it is allowed downstream.
Structured outputs validated with Pydantic, so a malformed agent response fails loudly instead of silently.
Every stage checkpoints — a failed step resumes from the last good state instead of restarting the run.

agent roles

Judge

scored outputs

Resume

from checkpoint

PythonLangGraphLLM-as-a-JudgePydanticCheckpointing

view on GitHub ↗

system architecture

9 components · 1 subsystem · 9 flows

02Multi-document RAG assistant

DocMind

Ask a question across a pile of PDFs and get one grounded answer.

A retrieval stack tuned end to end: chunking, reranking and query rewriting, measured against a baseline.

Ingests multiple PDFs, chunked at 1000 tokens and indexed in ChromaDB.
Retrieved candidates reranked with a Cross-Encoder instead of trusting raw vector similarity.
A LangGraph state machine rewrites weak queries before answering — 40% fewer hallucinations across 200+ test queries.
Beat a BM25 baseline on relevance by 28% on the same query set.
Shipped on Streamlit with sub-2s end-to-end latency on multi-PDF questions.

−40%

hallucination

+28%

vs BM25 relevance

<2s

end to end

LangGraphChromaDBCross-EncoderHuggingFaceGroqStreamlit

view on GitHub ↗

system architecture

10 components · 2 subsystems · 10 flows

03Agentic RAG on AWS Bedrock

Customer Care Agent

Support that remembers the conversation and calls tools when it needs to.

A production support agent on AWS Bedrock: retrieval, memory and tools behind one conversation.

Agentic RAG over 500+ FAQ entries using Amazon Titan Embeddings v2 and FAISS.
~92% retrieval accuracy on domain-specific questions.
Deployed on AWS Bedrock AgentCore with multi-turn memory that carries context across a full conversation.
Tool-calling workflows for the requests retrieval alone cannot answer.
CloudWatch observability on every invocation — average query resolution time down 35%.

92%

retrieval accuracy

−35%

resolution time

500+

FAQ entries

AWS Bedrock AgentCoreLangChainFAISSTitan Embeddings v2CloudWatch

view on GitHub ↗

system architecture

8 components · 1 subsystem · 8 flows

04Current work · Internovo Ventures · MumbaiUnder NDA

Agents for real estate

Two agents in production, a search harness underneath, a builder portal in progress.

Two agents in production, a traced search harness underneath, and a builder portal being built on top.

Architected and designed the WhatsApp lead-nurture agent across three phases, owning the backend design and implementation for all three.
Phase 1 — inbound: leads talk to the bot directly from the company's product website, where it qualifies intent and captures requirements in the conversation.
Phase 2 — outbound: the bot fires approved WhatsApp templates at leads as they land in the CRM, then holds a live conversation with whoever replies, writing every turn back to the lead record.
Phase 3 — handoff: qualified lead details are routed to the appointed salesperson, so a human takes over exactly at the point the lead is warm.
250+ leads moved from cold to warm with autonomous follow-ups.
A LangGraph property-search agent that reads portals, builder sites and the open web, then fills a fixed schema per project — possession, amenities, carpet area, RERA number, builder, location, nearby infrastructure and description — with every fact linked back to its source.
Both agents run on a search harness I built that mixes deterministic steps with tool-calling LLM stages, traced end to end with OpenTelemetry and Langfuse.
Currently building a builder-focused portal on top of it, starting with a competitor-analysis agent that maps a builder's rivals and turns the findings into an analytics dashboard.

250+

leads cold → warm

fields extracted

Traced

OTel + Langfuse

LangGraphWhatsApp APIPythonFastAPIOpenTelemetryLangfuseNext.jsClaude

source private

system architecture

13 components · 3 subsystems · 14 flows

mind

A polymath in progress.

Engineer by training, reader by habit, athlete by necessity. I am curious about how brains work, how people decide, and how to build machines that do both a little better. The graph on the right is the honest map.

First principles over pattern matching.

When something breaks I go down to the physics of the problem, not to the last Stack Overflow answer that looked similar.

Life rewards action.

Most of what I know came from shipping something too early and fixing it in public.

Measure, then talk.

Retrieval accuracy, latency, hallucination rate. If a number can be put on it, I put a number on it.

Read outside the field.

The brain solved memory, attention and retrieval long before we did. Neuroscience and psychology are where I steal my best ideas about agents.

The body is part of the stack.

Football, badminton, swimming, the gym. Thinking gets better when the machine running it is maintained.

Build

Agents · RAG · AWS · Full-stack. Hover a node.

notes

Writing is how I check my thinking.

view all notes

01draft4 min

First principles is a debugging tool, not a philosophy

Everyone quotes it. Almost nobody uses it at 2am when the retrieval pipeline returns garbage.

2026-09-01

02draft5 min

Working memory, context windows, and why my agents checkpoint

The brain has been running a resource-constrained agent for a few hundred thousand years. Some of its tricks port directly.

2026-08-24

03draft4 min

Notes on wanting to be a polymath in an age of specialists

Football, neuroscience, LangGraph, writing, swimming. The advice says pick one. Here is why I think the advice is slightly wrong.

2026-08-10

path

Where the action has been.

2022

HSC, Thakur College of Science and Commerce

Finished school in Mumbai. Started writing code that other people used.

2023

B.Tech, AI & Data Science at VCET

Vidyavardhini's College of Engineering & Technology. CGPA 9.02 so far, graduating 2027.

2024

Web Developer Intern, Acmegrade

Four full-stack apps on React, Express and MongoDB. Cut page load ~30% with query optimisation and code splitting.

2024 →

Freelance full-stack developer

Three-plus client sites on the MERN stack and AWS (S3, EC2, ECS Fargate). Backends handling 500+ concurrent requests.

2026 →

AI Engineer & Business Operations, Internovo Ventures

WhatsApp nurture agent, LangGraph property-search agent, a traced search harness (OpenTelemetry + Langfuse), and a builder portal with a competitor-analysis agent in progress.

2027

Graduate

Open to AI engineering roles where the agents have to work in production, not in a notebook.