H M Naimuzzaman Shuvo

A real-time AI voice tutor for Bangladeshi higher secondary students.

I architected and built end-to-end: a Python back end (FastAPI + Supabase) that orchestrates Pipecat bot subprocesses connected to LiveKit rooms for sub-second voice, with Gemini 3 Flash Preview as the LLM brain and Google Cloud TTS / STT for audio. The mobile app (React Native + Expo, NativeWind v4) renders a live whiteboard on which the bot draws via structured tool calls, with full pause-and-resume across sessions. The bot's LLMContext snapshots persist per turn, allowing seamless rehydration on resume with prior board state.

I designed a chapter-scoped lesson lifecycle and a 2-stage post-session analyzer that runs Gemini structured-output passes to (1) audit each session and (2) maintain rolling per-chapter mastery state used to personalize the next session. I also designed the entire UX—an Apple-Fitness-inspired system spanning Summary, Learn, Practice, and Notes tabs, with bilingual rules (Bangla content, English chrome) baked into a locked design token palette.

Built an AI agent platform that automates qualitative market research: founders draft an Ideal Customer Profile, the system recruits candidates from the open web via neural search, conducts voice interviews, and synthesizes findings across them with a strict, structured LLM output. The differentiator—a single tool registry exposed simultaneously over both a chat-completions HTTP API and a Model Context Protocol (MCP) server, so the same capabilities work from a custom web UI or any MCP client (Claude Code, Cursor) with zero duplication. I also designed device-flow authentication via URL-mode elicitation, lazy refresh-on-read for cross-interview synthesis, and client-agent ICP judgment to eliminate redundant LLM calls.

An all-in-one platform I built solo for live code collaboration and technical interviews. Real-time collaborative code editor, multi-language code execution (JavaScript, TypeScript, Python), WebRTC video and chat, live drawing canvas, and rich-text document editor—all in one room accessible via tokens, no login required. It has real-time sync over Socket.io, video via Agora, and code execution sandboxed in Node.js child processes.

An installable AI agent skill that turns any live web app into a narrated product walkthrough. The pipeline crawls the target with Browserbase and Stagehand, selects the most relevant states for a chosen persona using Gemini Flash vision, generates a voice-over with Gemini TTS, and composes an HTML preview and an MP4 with FFmpeg. I designed the BFS crawler with state-level deduplication, the structured-output selection pipeline, and the FastAPI orchestrator powering both the agent CLI and the web UI. One command returns JSON for downstream agents.