// work

Projects

Things I've built — AI systems, LLM apps, and full-stack products. Click any project for the full breakdown.

// kind words

What people say

Feedback from the teams and founders I've built AI systems with.

Junaid shipped our RAG assistant end-to-end — retrieval quality was the difference between a demo and something our team actually trusts.

AR A. RahmanProduct Lead, SaaS

Rare to find someone who can train the model and build the product around it. Clean APIs, real evals, no hand-waving.

SM S. MalikCTO, Fintech

The agent he built runs unattended every night. Guardrails, retries, tracing — it just works, and when it doesn't we know why.

JC J. ChenHead of Ops

Fast, communicative, and obsessed with latency and cost. Our inference bill dropped while quality went up.

MF M. FarooqFounder, AI Startup

Junaid shipped our RAG assistant end-to-end — retrieval quality was the difference between a demo and something our team actually trusts.

AR A. RahmanProduct Lead, SaaS

Rare to find someone who can train the model and build the product around it. Clean APIs, real evals, no hand-waving.

SM S. MalikCTO, Fintech

The agent he built runs unattended every night. Guardrails, retries, tracing — it just works, and when it doesn't we know why.

JC J. ChenHead of Ops

Fast, communicative, and obsessed with latency and cost. Our inference bill dropped while quality went up.

MF M. FarooqFounder, AI Startup

He treated our messy documents like a first-class problem. Chunking and reranking done right — answers finally cite the source.

LA L. AhmedData Manager

Delivered a full-stack LLM feature in weeks, not months. Streaming, function calling, the whole thing felt production-grade day one.

RI R. IqbalEngineering Manager

Great writing, greater engineering. His breakdowns of what actually works in production saved us weeks of trial and error.

DC D. CostaML Engineer

Reliable partner from prototype to deploy. Owns the system — data pipeline to the button a user clicks.

KY K. YusufVP Engineering

He treated our messy documents like a first-class problem. Chunking and reranking done right — answers finally cite the source.

LA L. AhmedData Manager

Delivered a full-stack LLM feature in weeks, not months. Streaming, function calling, the whole thing felt production-grade day one.

RI R. IqbalEngineering Manager

Great writing, greater engineering. His breakdowns of what actually works in production saved us weeks of trial and error.

DC D. CostaML Engineer

Reliable partner from prototype to deploy. Owns the system — data pipeline to the button a user clicks.

KY K. YusufVP Engineering
Multi-Agent AI Research Assistant screenshot

2026

Multi-Agent AI Research Assistant

Agentic RAG + web search over your PDFs & URLs

A full-stack, multi-agent AI research assistant that answers questions with citations from your PDFs and the web. Agentic RAG orchestrates retrieval, web search and reasoning across a scalable FastAPI backend, with a modern, responsive Next.js frontend.

Highlights

  • Agentic RAG + live web search with citations
  • Answers grounded in uploaded PDFs and URLs
  • Scalable FastAPI backend, Qdrant vector store
  • Modern responsive Next.js frontend

Built with

FastAPILangGraphQdrantNext.js
TrailGuard AI screenshot

2025

TrailGuard AI

Offline AI survival companion — Gemma on-device

When you're deep in the backcountry, signal disappears — but emergencies don't. TrailGuard AI puts a survival expert, trail navigator and emergency coordinator in your pocket: Gemma runs fully on-device via Google's LiteRT-LM engine inside a Flutter app, so everything works with zero internet connection.

Highlights

  • 100% offline — Gemma on-device via LiteRT-LM
  • Survival expert, trail navigator & emergency coordinator
  • Flutter Android app with downloadable APK
  • Built for the Google/Kaggle hackathon — video demo live

Built with

FlutterGemmaLiteRT-LMAndroid
RepoRadio screenshot

2025

RepoRadio

Your code is now on air — a voice AI host for any repo

RepoRadio turns any GitHub repository into a live radio show. Paste a repo URL and an AI host that has read the entire codebase explains it out loud — ask anything by voice, interrupt it mid-sentence (it stops when you speak), and ask for a roast when you dare. Groq's free tier plus a fully local Kokoro TTS voice means it costs $0 to run.

Highlights

  • Voice conversation grounded in the actual code
  • Barge-in — the host stops the moment you speak
  • Groq LLM + 100% local Kokoro TTS — $0 running cost
  • Roast mode, on request

Built with

PythonGroqKokoro TTSVoice AI
YouTube Content Producer screenshot

2025

YouTube Content Producer

Multi-agent system that automates YouTube content

A multi-agent AI system that automates the full YouTube content pipeline — topic research, trend analysis, script writing and SEO optimization — using OpenAI/LLM APIs and prompt orchestration. Real-time streaming workflows with FastAPI + Server-Sent Events, and JSON/PDF/CSV export. Written up as a Kaggle Agents Intensive capstone project.

Highlights

  • Agents for research, trends, scripting & SEO
  • Real-time streaming via FastAPI + SSE
  • JSON / PDF / CSV export pipelines
  • Kaggle Agents Intensive capstone write-up

Built with

PythonOpenAIFastAPISSE
Brain Tumor Detection Web App screenshot

2024

Brain Tumor Detection Web App

Deep-learning MRI classification in the browser

A medical-imaging web app that detects brain tumors from MRI scans. Upload a scan and a trained convolutional neural network classifies it instantly with confidence scores — a clean, accessible interface over a serious model, built as a screening-support tool.

Highlights

  • CNN-based classification of brain MRI scans
  • Instant predictions with confidence scores
  • Clean upload-and-result web interface
  • Trained & evaluated on public MRI datasets

Built with

PythonTensorFlowCNNComputer Vision
Medical Q&A Assistant screenshot

2024

Medical Q&A Assistant

RAG + LoRA fine-tuning for grounded medical answers

A production-ready medical Q&A assistant that gives efficient, evidence-grounded answers. TinyLlama fine-tuned with QLoRA for efficiency, FAISS-based RAG for fast evidence retrieval, and context-aware responses grounded in medical literature — served through a clean Gradio interface and a high-performance REST API.

Highlights

  • TinyLlama fine-tuned with QLoRA (efficient)
  • FAISS-based RAG for fast evidence search
  • Answers grounded in medical literature
  • REST API + clean Gradio web interface

Built with

TinyLlamaQLoRAFAISSGradio