Shubham Garje — Pune, IndiaAvailable part-time

Machine learning that
ships with an interface.

Computer vision, RAG pipelines and agent systems — and the frontend people actually use them through. I build the whole system, not the notebook.

Computer visionRAG & agentsProduct interfacesFrontend build
  • Best UICSI Hackathon 2026
  • 1.9M+ visitsMindSpark platform
  • Goldman SachsIndia Hackathon 2026
  • OSS mergedpygments · structlog · openapi-ts
  • Minor in ML & AICOEP Tech

Selected work

05
Full-stack

NETRA — Drone Surveillance & Response

A drone fleet control system: spot an incident from the air, decide which drone goes, and route it there.

YOLOv8 runs incident detection over the feed, then a weighted assignment picks the right drone — checking battery feasibility before committing — and A* routes it around no-fly zones. Operators watch it happen on a real-time fleet dashboard.

Role
End-to-end — model, backend, interface
Stack
Next.js, TypeScript, YOLOv8, React Leaflet, Three.js, PostgreSQL, SSE
Scope
Detection, assignment, routing, live fleet dashboard
Outcome
Best UI Award · CSI Hackathon 2026
Live siteRepo
ML

VaultSec — Document Intelligence for Financial Paperwork

Ask your bank statements and bills questions in plain English — with the arithmetic done by SQL, not by a language model.

Two retrieval paths, routed by intent. "What was that large transfer about?" is semantic, so it embeds the query and pulls the nearest chunks from a pgvector index. "Total deposits in July?" is arithmetic — which language models are unreliable at — so it runs a SQL aggregation over a transactions table built at ingest. Those transactions come from balance-delta classification: rather than trust a PDF's column layout, which OCR routinely mangles, it reads the running balance and infers each row's direction and amount from the change.

Role
End-to-end — retrieval design, backend, frontend
Stack
FastAPI, Python 3.11, PostgreSQL + pgvector (HNSW), React 18, TypeScript, Vite, Gemini/Groq, Fernet
Scope
Hybrid RAG + SQL routing, balance-delta extraction, encrypted at rest
Outcome
Live on Render + Vercel + Neon
Live site
ML

AI-Driven Exam Proctoring System

Detecting cheating material in exam halls, on-device.

Annotated over 1,500 images per class and trained a PyTorch CNN to 92% accuracy on chit detection, with YOLOv8 and ResNet50 handling detection and classification over the custom dataset. Training ran through reproducible pipelines on NVIDIA V100 DGX systems, and every run was judged on accuracy, inference latency and failure cases — the failure cases being the part that actually changed the models. Runs offline, so footage never leaves the room.

Role
Undergraduate Research Assistant, COEP · Jun – Dec 2025
Stack
Python, PyTorch, TensorFlow, YOLOv8, ResNet50, OpenCV, Roboflow
Scope
1,500+ images annotated per class; reproducible training pipelines on NVIDIA V100 DGX
Outcome
92% accuracy on chit detection, evaluated for latency and failure modes
Full-stack

COEP MindSpark Platform

Registration and event management for COEP's flagship technical festival.

Owned the platform end to end as Web & App Head — design through deployment — across the festival's registration and event-management flows.

Role
Web & App Head, Jun 2025 – present
Scope
Registration and event management, design through deployment
Outcome
1.9M+ site visits during the festival period
ML

AI-Powered Code Review Assistant

An agent that reviews Python changes with the surrounding context in hand.

LangGraph orchestrates the review: retrieve relevant context, analyse the change, return feedback.

Role
ML
Stack
Python, LangGraph, Gemini, RAG
Scope
Context-aware analysis, automated feedback

What I work with

model → interface

ML & Data

  • Deep learningPyTorch, TensorFlow
  • Computer visionYOLOv8, ResNet50, OpenCV
  • Retrieval & agentsRAG pipelines, LangGraph, LLM agents
  • DatasetsRoboflow, annotation, evaluation

Build

  • FrontendReact, Next.js, TypeScript, Tailwind
  • Backend & dataFastAPI, PostgreSQL, pgvector
  • Realtimeserver-sent events, live dashboards
  • AccessibilityWCAG AA, keyboard, reduced motion

Design

  • Design systemstokens, variants, documentation
  • Product interfacesflows, states, edge cases
  • Data & 3D viewsmaps, dashboards, Three.js
  • Prototypingclickable desktop + mobile

Merged upstream

Shipd by Datacurve · Apr–May 2026

Experience

COEP Technological University, Pune
  1. Web & App HeadCOEP MindSparkJun 2025 – present

    Own the registration and event-management platform for COEP's flagship technical festival, design through deployment.

  2. Undergraduate Research AssistantCOEP Technological UniversityJun – Dec 2025

    YOLOv8 and ResNet50 detection and classification on custom datasets, with reproducible training pipelines on NVIDIA V100 DGX systems.

  3. Full-stack InternEduNiaa

    Full-stack product work.

  4. ML & Full-stack InternAinoSoft

    Machine learning alongside full-stack delivery.

B.Tech, Metallurgy & Materials Technology, Minor in Machine Learning & AI — COEP Technological University, Pune, Sep 2024 – May 2028

Also built

github.com/ShubhamGarje99 ↗

About

Pune, India

I like problems where the model is only half the work. NETRA needed detection, routing, and a control room someone could actually operate under pressure. VaultSec needed retrieval that knows when to stop trusting the language model and hand the arithmetic to SQL instead. The seam between the two is usually the interesting part.

What shaped how I work most was debugging VaultSec. Five separate failures, and not one of them crashed or logged an error: embeddings that carried no semantic meaning, a vector column sized for a model I'd already swapped out, an encryption key that regenerated on every cold start. Every component reported success and the output was quietly worthless. I instrument the failure path first now.

I'm studying metallurgy at COEP with a minor in ML and AI. The materials side trained me to characterise something before trusting it, which turns out to transfer.

Have a problem that needs a model and an interface?

shubhamvijaygarje@gmail.comGitHubLinkedInRésumé