AI Engineer · Independent Researcher · LLM Agents & Computer Vision
I build AI applications and do research across AI, agents and computer science. I'm pursuing a BS in Computer Science (AI) at Sukkur IBA University, expected August 2027.
LinkedIn · Kaggle · Portfolio · Email
Self-Improving Harnesses Below the Capability Floor (2026)
Can a small model running on a laptop rewrite its own prompt, memory and tools and get better?
I re-implemented Continual Harness (Karten et al., 2026) for local Ollama models and tested it on
qwen3.5:0.8b and gemma4:e2b across 20 automatically checked computer tasks.
- Self-editing as-is made both models worse than a fixed harness (32% → 15% and 92% → 77%).
- Five added safeguards together (a larger refiner, a regression gate, retrieval memory, verified skills, structured errors) brought the small model back to baseline, with a suggestive gain on held-out tasks (27% → 60%).
- Reported with paired exact tests, Holm correction, per-seed results and a released dataset of 1,256 runs; one command rebuilds every table.
- trafilatura (7k★, ~13M downloads/month): fixed
target_languagediscarding every page for languages the detector cannot identify, such as Sindhi (#1113, merged).
- LecAssist — my agentic academic-assistance thesis (2026–2027), exploring lecture transcription, grounded question answering, and evaluation of multi-step agent workflows. The public thesis workspace includes an Aura desktop recording assistant and portal-automation work.
| Role | Organization | Dates |
|---|---|---|
| AI Engineer | Miami TV | December 2025–present |
| Junior AI Engineer | Memo Apps | September–November 2025 |
| Python Developer | Easy.co | June–August 2025 |
At Miami TV, my work includes agent workflows for research and content generation, a retrieval-based assistant, and streaming analytics. My previous roles covered educational software, workflow automation, and data-extraction pipelines.
| Project | Focus |
|---|---|
| Continual Harness on small models | Research code for the paper above: agent loop, sandbox, self-editing refiner, five safeguards, 20 tasks with checkers, and released data. |
| ProfDesk | Self-hosted assistant that puts Claude, ChatGPT and a private local model in one chat, usable from a phone. Built for teachers. |
| Tribunal | Multi-agent reasoning platform exploring structured debate, counterarguments, and synthesis with Python, TypeScript, Node.js, and LLM APIs. |
| Quantum-RAG | Python semantic-retrieval prototype with chunking, embeddings, vector search, and separate quantum-inspired experiments. |
| Rizzume | Resume builder connecting a React interface to Flask and LaTeX PDF generation. |
| CareToShare | Student study-material sharing with React, Express, MongoDB, Google OAuth, and Google Drive integration. |
Additional work includes ALIF-AI and a semantic book recommender.
ML & AI: PyTorch · TensorFlow · scikit-learn · Transformers · Ollama · NumPy · Pandas
Research: experiment design · paired significance testing · ablations · LaTeX
Software & deployment: Python · FastAPI · Flask · React · Node.js · Docker · Linux · GCP · MongoDB · Redis
For AI/ML engineering opportunities, research collaboration and project enquiries, contact me by email or LinkedIn.