Muhammad Ibrahim Salman

What I build

Most of my work is software that has to keep being right after the demo. UET GPT is an unofficial student project, not affiliated with UET Taxila. It retrieves the university's pages, with refusal instructions for empty retrieval and hedging for low scores. No answer-accuracy figure is published. B.L.A.S.T. reads scanned books and slide decks into clean text. The marketplace monitor I built during my internship stops and reports a failure instead of sending an empty digest. With Saif ur Rehman I built Nightshade, which applies AST-level rewrites intended to make scraped code less useful for model training. Compilation and regression tests check that the rewritten code still works.

I work mainly in Python, TypeScript and Java. Every project has a page on how it is built, and most of the code is on GitHub. Since July 2026 I've also been part-time co-founder of Corewell Systems, a small software agency building medical-billing tooling concepts against synthetic data (no client work is claimed).

How I work

I spend most of the time on the part demos skip: what happens when the input is bad, the source is down or the answer isn't in the documents.

I optimise for

Correct output, failures that are reported, and diffs a reviewer can read

Working defaults

Typed interfaces, tests before merge, and rate limits for external services

I avoid

Silent failures, tests that only cover the happy path, and configuration nobody can explain

What I measure, I publish

When I compare my work against another tool, the write-up includes where the other tool wins. On 14 hand-checked pages from one 1959 book, out-of-the-box Tesseract beat my OCR pipeline on 9 of them and on word error rate; the benchmark write-up has per-page scores and the script to run it again. A source-linked index of project measurements is on the evidence page.

You work with me directly, from the first call to the handover. I'm in Pakistan (UTC+5) and reply by email: ibrahim.pk848@gmail.com.