Experience
Research and engineering in parallel. The full version is in my CV.
Jun 2026 to nowNow
AI and Software Consultant · Multiple contracts
Remote
- Moved production document parsing from cloud APIs to a self-hosted MLX cluster, porting the Qianfan-VL OCR model and serving local LLMs, cutting the AI bill to near zero.
- Designed and shipped AI agents across production systems, from browser automation to document workflows, using Python, Temporal, Docker and Kubernetes on GCP.
- Audited the security of a production marketplace, fixing IDOR, broken access control and privilege escalation across its APIs and Stripe payment flows.
- Compressed a foundation model into a small edge model through pruning, retraining and quantization, deployed on Raspberry Pi for dashcams.
Jun 2024 to nowNow
Visiting Researcher · German Research Center for Artificial Intelligence (DFKI)
Kaiserslautern, Germany
- Post-trained a medical LLM for local deployment using PPO, DPO, RLHF, RLVR, PEFT, LoRA and quantization.
- Developed SLUM-i, a semi-supervised framework for mapping informal settlements from satellite imagery, with a newly annotated dataset.
- Researched self-supervised, semi-supervised and generative learning across SimCLR, MoCo, DINO, GANs, FixMatch-style methods and JEPA.
- Ran 14,000+ GPU-hours on up to 115 concurrent GPUs (NVIDIA B200, H200, H100, A100, RTX A6000).
Nov 2025 to May 2026
Research Associate · National Center for Artificial Intelligence (NCAI)
Islamabad, Pakistan
- Worked on E-TCAV, a framework for efficiently approximating TCAV scores with penultimate-layer proxies, evaluated across architectures and datasets.
- In a personal project alongside the role, studied adversarial neural cellular automata, where independently trained networks compete on a shared grid, explaining their bistable dynamics through mean-field analysis and targeted ablations.
- Co-supervised two MS theses as a BS graduate, on periocular biometrics and NSCLC immunotherapy response.
Jan 2025 to May 2025
Backend Engineer · Vyro.ai
Islamabad, Pakistan
- Developed the ImagineArt developer platform (platform.imagine.art) and its public generative AI APIs.
- Built low-latency LLM and generative media streaming with goroutines for chatlyai.app and imagine.art.
- Refactored the LLM backend around dependency injection, and optimized ChatlyAI's slow Cassandra queries.
Apr 2022 to Jan 2025
Freelancer · Fiverr and direct clients
Remote
- Delivered 50+ Next.js, MERN and Python projects, including e-commerce stores, admin dashboards and client websites.
- Built and deployed full-stack products with payments and cloud storage across Vercel, AWS, Cloud Run, Firebase and Hetzner.
- Sped up PostgreSQL by 25x through indexing on a system handling 10,000+ daily transactions, and helped mitigate a production DDoS attack.
Mar 2024 to Sep 2024
Open-Source Contributor, Google Summer of Code · CVAT
Remote
- Designed a configurable keymap system for CVAT's React and TypeScript annotation platform, with scope- and prefix-aware conflict detection and persistent user preferences.
- Re-architected shortcut registration into modules across the annotation interface, with Cypress end-to-end tests and documentation.
- Fixed warped frames from rotated videos and added a shortcut for pinned properties before the program; 8 PRs merged in total.
Jan 2024 to Jun 2024
Software Engineer · Thinksoft
Remote
- Co-developed WebbsAI, a generative AI platform that turns prompts into complete websites.
- Built product features across the stack with TypeScript, Next.js, Zustand, Firebase, MongoDB and AWS S3, supporting concurrent generation workflows.
Jun 2023 to Jan 2024
Software Engineer · OneScreen
Islamabad, Pakistan
- Built the React frontend of an LLM presentation maker that generates complete slide decks from prompts.
- Prototyped LLM agents and RAG pipelines with Python and LangChain, using BERT embeddings, MongoDB vector search and reranking.
- Implemented Redis-backed leaky-bucket rate limiting for LLM API calls.
Education
Nov 2021 to Jun 2025
BS Computer Science · National University of Sciences and Technology (NUST)
Islamabad, Pakistan
- Thesis at DFKI: Semi-Supervised Learning for Slum Segmentation using Remote Sensing and Deep Learning. Advisor Dr. Muhammad Imran Malik, co-advisor Dr. Muhammad Naseer Bajwa.
- Coursework included reinforcement learning, artificial intelligence, digital image processing, linear algebra, and probability and statistics.