PakGPT
FeaturedAutonomous voice AI — record → transcribe → LLM + tools → TTS. 5,000+ Play Store installs, 4.9 rating in 3 weeks.
AI Engineer · Lahore
I build LLM systems, agents, and production AI pipelines.
AI Engineer with 5+ years of production Python and 3+ years shipping LLM-powered systems end-to-end. At Xref I built an AWS Bedrock insights engine for 10,000+ daily users, shipped an MCP server for agent SQL access, and migrated a 10-year Django monolith to microservices. Solo: PakGPT (5,000+ installs), Uplix (4-agent LangChain pipeline), Vidora, and Lifeflow OS — a personal job-search operating system.
About
I'm Muhammad Abdullah, an AI Engineer based in Lahore, Pakistan. I focus on LLM pipelines, multi-agent orchestration, and the infrastructure that makes them reliable in production — not demos.
At Xref I work on Bedrock-powered insights, MCP tooling for agents, and microservices at scale. Outside work I ship solo products and run Lifeflow OS — my personal job-search command center.
Stack lately

Featured
Personal job-search operating system — manual queue with Q&A from Supabase, interview kanban with rejection learning, multi-provider LLM chat with job-search context, and ApplyPilot automation behind it.
Experience
Aug 2023 – Present
Projects
Three solo AI products plus open source. More on GitHub.
Autonomous voice AI — record → transcribe → LLM + tools → TTS. 5,000+ Play Store installs, 4.9 rating in 3 weeks.
4-agent LangChain product photography pipeline — orchestrator delegates validation, prompt engineering, generation, and QA workers.
Chained AI video translation: Whisper → OpenAI → TTS. Celery task graph with validation between stages. 30+ videos/day.
Also built
1,000+ GitHub stars · 50k+ monthly PyPI downloads. RAG layer + MCP server for live news in agents.
Serverless TTS — 300+ voices, 120+ languages on AWS Lambda.
Autonomous news → Instagram agent with LangChain ReAct loop.
Education
Bachelor of Computer Science · CGPA 3.35
Dean's List 2020 (top 10%)
Capstone: Urdu handwriting recognition with CNNs — 94% accuracy on 175k+ samples.