SleepWellSystem
● CurrentBuilding multi-agent reasoning systems with a Google research lead. Scalable LLM integrations for real-world workflows.
Fresh BS in Artificial Intelligence from FAST NUCES with 2+ years shipping real models. Currently an L2 AI Software Engineer collaborating with a Google research lead on multi-agent architectures. Published at IEEE INMIC 2.0, 2nd at TeknoFest 2024.
Three threads, one question: how far can a small team push AI agents that actually reason?
Partnering with a research lead at Google on reasoning and workflow automation for autonomous agents. Designing LLM integrations that hold up outside benchmark conditions.
Water-quality assessment pipeline: VLM extracts context from water imagery, an LLM writes the environmental report. Shipped end-to-end on Hugging Face + FastAPI.
Optimizing model performance and scalability across large Intel-provided datasets. Quietly one of the more educational things I do.
From annotation work at Scale AI to research labs to contract agent engineering — I've seen AI from every angle of the stack.
Building multi-agent reasoning systems with a Google research lead. Scalable LLM integrations for real-world workflows.
CV-based football analysis system with real-time object detection & tracking. Also: an NLP calling bot with speech recognition and dialogue.
Fine-tuning LLaMA and image-to-text models on Intel datasets, optimizing for scalability and downstream deployment.
Lifted model accuracy 60% → 93% with YOLO, ResNet, Inception. Cut video generation from 2 days to 10 minutes. Built the Node/Express API around it.
UAV modeling with LSTM, brain-therapy CNN analysis, satellite image dehazing, GAN-based recommenders. Two research codebases on GitHub.
Where the discipline for clean training data came from. Senior-tier review and annotation for ML pipelines.
Automated video + audio editing workflows with MovieLip and pydub. First taste of end-to-end media pipelines.
Agent systems, VLM/LLM product work, applied research. Pakistan-based, fully remote, full-time.
Four featured projects. Each one shipped, each one taught me something I still use.
A water-quality assessment pipeline that turns a photo of contaminated water into a readable environmental report. A vision-language model extracts visual evidence, a large language model writes the analysis, and both get packaged into an end-to-end system deployed on Hugging Face and FastAPI.
Also won Ideathon UBIT. Published the comparative analysis between VLMs and LLMs for climate monitoring at IEEE INMIC 2.0.
> query: "explain GAN collapse" > reformulated: "Describe mode collapse in generative adversarial networks, with 2 mitigation strategies." > retrieved: 6 docs > latency: 340ms
An auto-prompting layer that rewrites lazy user queries into prompts that actually work. Built on Hugging Face + LangChain + LangSmith + ChromaDB, with a Flask backend and a web UI. Removed the prompt-engineering step for researchers reading across dozens of papers.
Four models compared on implicit feedback — BPR, APR, CollaGAN, ACAE — benchmarked across Precision, Recall, HR, NDCG, MAP, and F1. Full adversarial training loop, evaluation plots auto-exported to PDF. My introduction to reading a recsys paper and rebuilding it from scratch.
Five small programs that each killed a recurring task: WhatsApp bot, email automation, file organizer, video editing, certificate generator. Together they cut repetitive work by up to 80% for the teams using them.
Filled dots are the ones I use weekly. The rest I'm fluent in but don't write love letters about.
Looking for a remote AI engineer who has read the papers and also shipped the thing? I'd love to hear what you're working on.