AV Abhay Vaishnav
Taking freelance projects Ahmedabad, India IST

अभय वैष्णव

I build AI that has to work on live data — not just in a demo.

Full-stack & AI engineer. Right now that means computer-vision pipelines that read kill feeds off live esports streams, and chatbots that answer from a company's own documents. I take freelance projects end to end: the model, the backend, the dashboard, the deploy.

Start a project Résumé (PDF)
Now
Full-Stack & AI Engineer at Bluexkye Software, on 16Score & Meta Ninja
Good at
LLM apps & RAG, YOLO + OCR pipelines, FastAPI / gRPC backends
Before
Conversational-AI backend for Ondewo, an Austrian startup
Works with
Startups, agencies and teams in India & abroad
● Rec · Cam 01
Slow-mo 0.3×
Frame 000000
0 tracked

A toy tracker. The kind of thing I build, in miniature. Hover to get detected · click for slow-mo Tap to toggle slow-mo

01

Selected work

Three systems I built in 2026. The constraint in all of them: the data is messy, it arrives live, and the output has to be right.

Live match stats, read straight off the broadcast

A real-time pipeline that pulls kills, deaths and assists out of mobile battle-royale and MOBA streams (Free Fire, BGMI, Mobile Legends) while the match is still being played.

  1. OBSstream
  2. Captureframes
  3. gRPCtransport
  4. YOLO + OCRdetect & read
  5. TMStournament system
  6. ViewersWebSocket push
  • Microservices backend: Next.js viewer app, React + Vite tournament and ops admin, FastAPI with internal gRPC, a Kafka-fed AI pipeline and WebSocket live push.
  • Also designed a tier-based snake-draft fantasy tournament mode for BGMI.

Fixing a pipeline that missed most of the kills

63%+of kill events were being missed before the redesign

Detection and OCR were tangled together, so one bad read meant a lost kill. I pulled them apart.

  1. 01Every killfeed row gets a sequence number the moment YOLO sees it
  2. 02Perceptual-hash fingerprints give each row an identity, so duplicates collapse
  3. 03The Hungarian algorithm assigns rows when several kills land at once
  4. 04ByteTrack-style retention keeps rows alive between frames
  5. 05Unreadable victim name? OCR → fuzzy roster match → knock history, run by a small deterministic state machine

Recall went up sharply, and every recovered name is traceable to the rule that produced it.

Catching edited tournament screenshots

Some players edit their kill/death/assist numbers before submitting result screenshots. This system checks each image through independent lenses and only flags it when they agree.

An evidence-fusion engine (levels L0–L6) needs multi-layer consensus per verdict. That keeps false positives down, which matters when the outcome is disqualifying someone.

Also built

  • Email MCP server — lets AI agents draft, send and triage email as tools.
  • Production chatbots on LangChain with intent routing, session memory and knowledge-base grounding.
  • RAG pipelines on RAGFlow for document-aware Q&A over internal knowledge bases.
  • Ondewo conversational AI — NLP intent & entity extraction, gRPC/protobuf services, CRM integrations and React admin dashboards for a call-center automation platform.
02

What you can hire me for

Fixed-scope projects or ongoing contract work. If it's not on this list, ask anyway. I'll tell you straight if I'm not the right fit.

  1. 2.1

    Chatbots & RAG over your documents

    Support bots and internal assistants that answer from your own knowledge base, remember the conversation and route intents properly, instead of making things up.

    LangChain · RAGFlow · OpenAI / Claude / open models

  2. 2.2

    Computer vision on video & images

    Detection, OCR and tracking for live streams, screenshots and scanned documents, plus the unglamorous dedup and fallback logic that makes the output trustworthy.

    YOLO · PaddleOCR · ByteTrack · PyTorch

  3. 2.3

    AI agents & MCP servers

    Give an LLM agent real tools: your inbox, CRM or internal APIs, with sensible limits on what it's allowed to do.

    Model Context Protocol · Python · tool design

  4. 2.4

    Backends, APIs & dashboards

    FastAPI or Spring Boot services, gRPC between them, Postgres and Redis underneath, Docker and CI/CD to ship. A React admin panel on top when you need one.

    FastAPI · Spring Boot · gRPC · PostgreSQL · React

03

How a project runs

  1. Step 1

    A short call

    You explain the problem. I ask awkward questions about your data and tell you honestly what's realistic.

  2. Step 2

    Scope on one page

    What gets built, what doesn't, timeline and price, agreed in writing before any code.

  3. Step 3

    Weekly demos

    Working software on a link you can click, every week. No status reports instead of progress.

  4. Step 4

    Clean handover

    Code in your repo, deployment notes, and a walkthrough so your team isn't stuck depending on me.

04

Toolbox & background

AI / ML
LLMs, LangChain, RAGFlow, RAG pipelines, MCP servers, NLP, prompt engineering, PyTorch, TensorFlow, YOLO, OCR
Backend
Python / FastAPI, Spring Boot, REST, gRPC & Protobuf, Kafka, WebSocket, event-driven microservices
Frontend
React, Next.js, Vite, TypeScript, Tailwind CSS
Data & infra
PostgreSQL, MySQL, Redis, Docker, CI/CD, Linux, Git
Testing
Pytest, JUnit, integration testing, clean architecture
Languages
Python, TypeScript, JavaScript, Java

Experience

  1. Dec 2025 — now Full-Stack Developer & AI Engineer Bluexkye Software, Ahmedabad
    • Apr 2026 — now 16Score & Meta Ninja, in-house AI and esports products
    • Dec 2025 — Mar 2026 Ondewo (Austria), client engagement on conversational AI

Education

  1. 2022 — 2026 B.Tech, Computer Engineering Ganpat University, Gujarat

Wins

  1. 2024 · 2023 Winner & runner-up, Convergence Inter-college technical competition, team “The Dominator”
  2. Hackathons GeeksforGeeks × Accenture “Hack the Future”, UIDAI (Aadhaar)

05 — Contact

Got something that needs building?

Email me a few lines: what it should do, a rough timeline, and a budget range if you have one. I'll reply within a day.

vaishnavabhay536@gmail.com