What I've built

Projects.

2026 Featured

Cutting AI diagram costs to zero (Autonise intern work)

Problem

Ask an LLM to draw an SVG and the geometry falls apart. It has no sense of space or coordinates. Doing it through commercial APIs costs ₹10 to ₹30 a diagram.

Approach

The model writes code, not pictures. A small Matplotlib DSL for diagrams, SMILES for molecules, rendered by local compilers. DocLayout-YOLO finds regions in exam papers. I moved inference to a self-hosted Qwen 35B.

Impact
  • Broken geometry gone. The compiler can't emit what it can't place
  • Per-diagram cost down to ₹0 on our own GPUs
  • Diagram detector scores 96.7% on the test set
  • AI judges contradicted themselves, A beats B beats C beats A. That is why round-robin won
AutoniseLLM SystemsProgrammatic IRDocLayout-YOLORDKitvLLMEvaluation
2026

Article recommendations from reading time (Info Me)

Problem

Almost no behavior data per user, and generic embeddings blur together in search results.

Approach

Weight interests by dwell time, pull top-5 by cosine similarity, rerank with two small Llama 3B agents on a local GPU. Then I checked my own metric. Found it broken.

Impact
  • Recommendations weighted by actual reading time
  • Two stages: embedding search, then a local model reranks
  • Caught my own evaluation metric failing on blurry categories. Wrote it up instead of hiding it
Info MeRecommendation SystemsLlama 3BDual-AgentTelemetryEvaluation
2025 Featured

Study assistant that turns playlists into quizzes

Problem

You can't study from a three-hour playlist. Transcripts run longer than the model can read, and the good parts stay buried.

Approach

An assistant on Azure that pulls in the playlist, compresses transcripts in stages, then answers questions and makes notes, flashcards, and quizzes. If one provider dies it fails over. Repeat questions use cached answers. One bad video can't sink the playlist.

Impact
  • Live and used as a study tool at yt-helper.shubhsareen.com
  • Ask questions of multi-hour playlists
  • Keeps working through provider outages and rate limits
LLM PipelineRAGAzure Container AppsOpenRouterEmbeddings
Source Live App
2025

App that writes captions for photos

Problem

Describing a picture in words takes vision and language working together. Classification alone does not cut it.

Approach

A captioning pipeline built from scratch in IIT Bombay's Seasons of Code. ResNet-50 encoder with transfer learning, attention-based LSTM decoder, GloVe embeddings, 500K+ training tokens.

Impact
  • A full pipeline trained all the way through
  • Attention weights show which pixels each word came from
  • All training and inference code on GitHub
PyTorchResNet-50Attention LSTMGloVe
Source
2026

Experiment: splitting tasks across small AI agents

Problem

One big prompt falls over when a task needs splitting up, handing off, and remembering things across steps.

Approach

Small agents for splitting tasks, routing, and answering, wired into a graph that passes state between steps.

Impact
  • Multi-step work split across specialist agents
  • Conversations with memory you can steer
  • An honest log of where the agents failed and why
LangGraphMulti-AgentGoogle ADK
Source
2025

Campus-Scale Sports & Operations Portal

Problem

13,000+ students, no central place for fixtures or schedules. Just scattered group chats.

Approach

I built the official portal as Web Convenor. React frontend, Django backend, fast and readable on phones.

Impact
  • The whole campus uses it
  • ~30% faster loads than what it replaced
  • One place for events, schedules, contacts
ReactDjangoResponsive
Source