My Projects

An agentic system that goes beyond prompt-to-code generation — it inspects a repository, understands how components fit together, reviews implementation quality, writes and runs tests, and iterates on fixes. Built to behave like an autonomous software engineer, not just an assistant.

A terminal coding assistant powered by AWS Bedrock — the same interactive-agent feel as Claude Code, but running across 47+ models spanning Anthropic, Qwen, DeepSeek, Llama, and more. Ships with an SSO-aware setup wizard and a one-line installer.

A full-stack monitoring system — FastAPI backend, React dashboard — that tracks service health on a schedule and only alerts on real problems, not single-blip noise. Slack and email notifications, JWT auth, and response-time analytics.

A local-first, production-style fraud detection simulator — Kafka event streaming across independent FastAPI/consumer services, a two-stage rules-then-ML detection pipeline, and a live SSE dashboard for alerts.

A tool that streamlines VPS provisioning — turning a normally fiddly multi-step server setup into a few clicks. Backend infrastructure tooling built to make deployment less painful for solo projects.

An MCP server that exposes a personal notes tool to any MCP-compatible client, letting an AI assistant read and write developer notes directly. A small, focused piece of applied MCP plumbing.

Every file here is code written while completing NeetCode's ML course — progressing from gradient-descent fundamentals through attention, transformer blocks, KV-caching, and grouped-query attention to a working GPT. First-principles, not a wrapper around someone else's model.

Generates full podcast episodes from a single topic — scripting, multiple speakers, and a Streamlit UI to drive it end-to-end. A compact multimodal GenAI app people can actually sit down and use.

An open-source infrastructure tool for LLM applications. Deterministic prompt versioning, A/B testing across models, and evaluation metrics — bringing DevOps-style rigour to the LLM development lifecycle. A backend-first take on managing prompts like code.

A lightweight gateway that unifies access to multiple LLM providers behind a single API, so a backend can route between or swap models without rewiring. A practical piece of LLM plumbing built to sit in the request path of real services.

A fine-tuned Gemma 4 (E4B) model specialized in system design — distributed systems, scalability, API design, database selection, caching, and backend architecture patterns. Trained on a synthetically generated, critic-filtered dataset of structured system-design Q&A.

A voice-to-voice conversational assistant built on Amazon Nova Sonic that answers from your own documents via a RAG pipeline. Speech in, grounded speech out — a full applied-GenAI loop wired up as a deployable backend service.

A near-realtime voice chatbot that understands and responds across multiple Indian languages, built on the Sarvam speech and language stack. Focused on low-latency turn-taking for a natural, spoken conversational experience.

A playful generator that turns a prompt into a finished meme using Google's Gemini and Imagen models — Gemini for the caption and concept, Imagen for the image. A small end-to-end multimodal GenAI app that people actually use.

An anime recommendation system that uses an LLM for semantic understanding of taste and titles, wrapped in an LLMOps setup for tracking and iteration. A hands-on take on recommenders in the GenAI era, from data to serving.

A compact Todo application in Python/Flask, packaged and deployed with docker-compose. Small on purpose — it's the end-to-end deployment story in miniature: build the service, containerize it, and ship it on a plain server.