Sabarish Raja Ramesh Raja

Production-grade React Native app with Supabase backend, distributed as a standalone Android APK via EAS Build — integrating multi-model Claude API (Sonnet 4.5 vision + Haiku 4.5 chat) for clothing classification and conversational outfit generation.

Claude runs behind a Supabase Edge Function acting as a secure AI proxy — the mobile bundle never touches the API key — turning image uploads into schema-validated relational data with row-level security. Multi-stage outfit pipeline: natural language → Postgres filtering by formality/season → Claude visual reasoning over wardrobe images → ranked combinations with gap analysis and affiliate links. Handles Indian ethnic wear (kurtas, sherwanis, dhotis) as first-class categories. Voice input via Whisper, weather-aware prompts, PostHog + Sentry observability.

• React Native
• Supabase
• Claude API
• Whisper
• EAS Build

A two-part observability suite for AI coding agents — remember what happened, verify what's claimed. Hindsight indexes Claude Code's session transcripts into local SQLite and surfaces where you left off via a terminal briefing, a web dashboard with a living architecture doc, and a statusline.

Oversight hooks the Stop event to re-run tests, re-check git and the filesystem, and block false "done" claims before they reach you. Both ship from a plugin marketplace and run 100% locally.

• TypeScript
• Python
• SQLite
• Claude Code Hooks

A full-stack web app for writing letters to your future self or others, held and delivered on a date you choose. React + TypeScript front end backed by a Node/Express API and a Supabase (Postgres) datastore with row-level security.

Server-side rendering pipeline uses headless Puppeteer to turn each letter into a polished, print-ready PDF on demand. Deployed on Vercel with a decoupled API service.

• React
• TypeScript
• Node / Express
• Supabase
• Puppeteer

Architected a full-stack LLM agentic platform using LangGraph and GPT-4 Vision that resolved 70%+ of customer support and fraud cases autonomously.

Designed semantic search using vector embeddings to improve book discovery; implemented JWT-secured RESTful APIs and an admin dashboard with real-time analytics for AI-driven case management.

• LangGraph
• GPT-4 Vision
• Vector DB
• JWT

Ensemble of Isolation Forest + XGBoost + SVM achieving 97.0% PR-score and 96.3% ROC-AUC on a highly imbalanced fraud dataset using ADASYN and NearMiss resampling.

Findings published in AIP Conference Proceedings; presented at ICRAIC 2024.

• XGBoost
• SVM
• ADASYN
• Imbalanced data