micro1
Sep 2026 – Present
LLM Red Teamer
Tests frontier models and agents for jailbreaks, prompt injection, tool misuse and information disclosure. Each finding is written up as a reproducible case.
Ali Basheer
LLM red teaming and AI data quality.
Work
Cohere
Dec 2024 – Jul 2026
Senior - Data Quality (STEM)2025 – Jul 2026
Data Quality (STEM)Dec 2024
Data quality on STEM training and evaluation data. Criteria agreed with dataset owners, audits in Python and SQL on BigQuery, and pipeline QA. Promoted to Senior within two months.
Consulting
2020 – 2024
Oxy · Veolia · Bilfinger
Environmental, safety and operations consulting for oil and gas, industrial and waste-management clients.
Craftpine
Founder · since April 2026
A research and development studio in Toronto. Every program is published in full on craftpine.com, including the ones that closed.
Corner
Payments and inventory for Toronto corner stores. Independent corner stores often run on a single register, so Corner started from that till. It was built with two AI assistants, took one live charge, then closed on the numbers.
- Build window
- 76 days
- Commits
- 173
- API routes
- 57
- Tables
- 20, row-level security on every one
- Test suites
- 44
- Adversarial
- 19 scenarios and 71 checks, all held
- Live charge
- $1.99 on Amex, July 9, refunded
A 3B language model and speech-to-text ran together on a $250 Jetson-class device, 2.6 seconds end to end. Grammar-constrained decoding cut a reply from 8.8 seconds to 2.9.
How it worked and why it closed520 Barton
Could an unattended store with no staff pay for itself on one Hamilton corner? Four ideas were tested and none passed. The best published throughput for an unattended fresh-drink machine is 6.8 cups a day, and the smallest version of the plan needed 12.
The study, trimmed to one pageTerrasope
Mineral sensing for ground with no GPS, no satellite link and no people. Only the record layer is built. Entry 1 ran on September 10 and missed, and hardware work stopped there.
Where it standsDieselgate
Best Seminar · University of Toronto · 2019
Defeat software satisfied the laboratory cycle and failed in ordinary driving. Manufacturers repeated that method through the Dieselgate emissions scandals. Cost-of-life calculations priced the surplus NOx once dispersal models had moved it from the tailpipe into the street. Algorithms identified whether the engine was on the test cycle or on the road.