Muhammad Jahanzaib Awan

How I work

A model that scores well for the wrong reason is worse than one that scores honestly, because it fails silently in production. So before I trust a result, I try to break it.

On a UAV object-classification project, a naïve random train/test split gave a confidently high macro-F1. But consecutive frames of the same vehicle are near-duplicates, so random splitting leaked the test set into training. A track-aware split, keeping each object on one side of the divide, dropped F1 by roughly 0.5. That collapse is the real story: it's the difference between a number you can put in a report and one that would embarrass you in deployment.

The same discipline runs through everything here: strong baselines before big models (a tuned GRU beating a LoRA-fine-tuned GPT-2 on the Cornell Movie-Dialogs Corpus), the right metric for the data, and reproducibility, fixed seeds, logged configs, and runs tracked in Weights & Biases so every figure on this site can be regenerated.

Experience

Automated production system, independent project

Founder and sole engineer · June 2026 to present · Python, GitHub Actions, Cloudflare Workers, FFmpeg, React, REST APIs

I designed, built and run an unattended content production system that has published autonomously twice a day for months. It generates each piece from a topic queue, assembles media, renders video and distributes to four platforms without anyone touching it.

The interesting part is not the pipeline, it is what keeping it alive has taught me. It runs a multi-provider model and speech layer with automatic failover, because a quota exhaustion or a lapsed card at one vendor should degrade the output, not stop production. It manages OAuth credentials across five external APIs, with scope auditing and expiry alerting added after token expiry caused repeated outages. And it gates its own output: every render is measured for transcription error against the source script, speech pace, loudness and frame-level visual completeness, and regenerated until it passes.

My favourite bug from it is a lesson in the same thing this whole site is about. Videos were shipping with blank frames over live audio, and every signal was green: the receipts said all assets were staged, the black-frame detector said the picture was fine, the analytics said retention was healthy. All true, and all blind. I pulled frames out of a published file, found five seconds of empty screen, traced it to a token-alignment defect that placed the opening captions on top of the wrong audio, fixed it with regression tests, and added the automated visual check that now refuses to publish a render whose frame is empty. A passing status check is not the same thing as a working system.

Bank Residential Support Worker, Platform Childcare Ltd

Leicester, UK · December 2025 to present · Part-time

Specialist 1:1 support for children and young people with complex needs, including autism, learning disabilities, ADHD and physical disabilities, across residential, educational and community settings. I implement individual support plans, maintain daily records and incident reports to safeguarding standards where accuracy is formally accountable, and adapt how I communicate between a child, their family, colleagues and external professionals. Worked alongside full-time postgraduate study.

Education

Taught modules span machine learning, neural systems, NLP, computer vision, robotics, evolutionary computation, fuzzy logic and research methods. Most of the case studies on this site grew out of that coursework.

What I work with

What I'm looking for

Graduate Machine Learning Engineer and Data Scientist roles where evaluation actually matters, teams that care whether a result is real. I'm Leicester-based and open to remote or relocation, available from October 2026. I hold the right to work in the UK on the Graduate visa route after my MSc completes.