Selected Work6 projects

Selected Work

Systems built under real constraints, a deadline, a jury, or a production pipeline that had to keep running after I left the room.

AI & Maritime Safety

Seaguard

Code

Problem

Fishermen navigate open water with no reliable way to know where the fish are, or when the weather is about to turn against them.

Solution

A mobile assistant for fishermen that detects potential fishing zones, pushes safety and SOS alerts, and adds catch logging, trip planning, GPS navigation, weather updates, and multi-language access on top.

Architecture

React Native app talking to a FastAPI backend; a TensorFlow model scores potential fishing zones from satellite and weather signals, with Supabase handling accounts and trip data.

React NativePythonFastAPILLMsSupabaseTensorFlow

Most Product Readiness, Hackcelestial 2.0

AI & Cybersecurity

Kavach.ai

Code

Problem

Analysts need to know what a suspicious file actually does before it's safe to let it near production, and most tools only tell you after the fact.

Solution

A desktop cybersecurity tool that detonates suspicious files in an isolated Docker sandbox, scores them with a trained ML model, and explains the verdict instead of just flagging a hash.

Architecture

An XGBoost model trained on malware samples handles detection, a file-entropy check catches obfuscated payloads, and a Docker sandbox captures system-call behavior while LLaMA 3.2 interprets what went wrong, all wrapped in an Electron desktop app. Tested live against GajShield's own malware samples in front of the judges.

ElectronPythonDockerFlaskXGBoostLLMs

1st of 2,000+ teams at GajShield, KJSCE

AI & Finance

FinSaathi

Live Code

Problem

Most people don't have a financial advisor, and the government schemes they already qualify for go unclaimed because no one told them.

Solution

An AI-powered platform combining stock analysis, trading simulation, financial-document processing, and a chatbot for scheme discovery into one place.

Architecture

React frontend on a Node/Express backend; the Agentic Document Extraction API pulls structured data out of financial documents, and Twilio pushes real-time stock updates straight to WhatsApp.

Next.jsMongoDBPythonFlaskPyTorchLLMs

Live · 1st of 2,000+ teams at Datathon, KJSCE

AI & Finance

LoanSaathi

Code

Problem

Over 40% of loan applications in India get rejected on paperwork errors, and predatory hidden terms exploit applicants who can't read the fine print.

Solution

A multimodal AI loan assistant that reads and explains loan documents, flags harmful clauses, and walks users through the entire application in 8 Indian languages.

Architecture

Mistral Saba 24B plus a fine-tuned FinLLaMa handle document understanding, PaddleOCR extracts the paperwork, Sarvam AI drives agentic parsing and conversation, and Supabase stores session data with nothing retained after processing.

PythonFastAPINext.jsMongoDB

1st of 5,000+ teams at The Great Bengaluru Hackathon

Web Platform

SwapIt

Code

Problem

Students overpay for study material that's used for one semester and then thrown away.

Solution

A platform for affordable study-material exchange between students, using Google Sheets and Drive as a lightweight, zero-cost data layer.

Architecture

Next.js frontend authenticated through Google OAuth, with Sheets and Drive APIs standing in for a database and file store.

Next.jsGoogle APIsOAuthGoogle Sheets

Shipped

AI & Misinformation

FactFinder

Code

Problem

Misinformation spreads faster than anyone can fact-check it by hand.

Solution

A fake-news detection platform that scores the authenticity of a news article using an ensemble of transformer and classical models.

Architecture

A BERT-based classifier ensembled with traditional models, trained on the 72,000-article WELFake dataset and served through a Flask backend.

FlaskPythonLLMsTransformers

2nd of Hackanova 3.0, TCET (36hr)