Work
My own products, one built with a team, and audits of code that AI wrote. Each entry says what I did and links to the proof.
Showing 6 of 6
RentPilot · Solo build, 2026
A car-rental marketplace, built by one engineer and an agent team.
Live at rentit.mobi. About 1,200 commits and over 200 merged pull requests, each one through a review gate I own.
Evidence: git history, checked 27 Sep 2026Read the case study



Ta'beer · Two-person studio, beta
A dream-interpretation app whose model may not invent a source.
Arabic and Sorani, dictionary first. Retrieval is deterministic, the model gets one call, and code rejects any answer that cites what it didn't retrieve.
Evidence: 200-case eval set, repo docsCase study after launch

RedesignAI · ALB Group team, live
An Instagram handle becomes a website; the model writes a plan, never code.
The model picks from a closed catalogue of patterns and a compiler writes the HTML. A model switch took one benchmark job from 188 s to 46 s, and first-try valid output from 71% to 100%.
My part: senior engineer and product owner. I rebuilt the generation pipeline and the editor, added the animated progress steps, and built the design kit and templates, all with agentic workflows.
Evidence: model benchmark in the repo docsredesignai.coWith ALB Group alb.az
Audits · Client work, anonymised, 2026
Two audits of AI-built apps, and what the AI got wrong.
A Next.js and Supabase app with AI-written summaries: recon, verified findings, a security review and a fix estimate. A payroll SaaS whose founder builds with AI: architecture review, then the fixes.
Evidence: finding classes and method; the reports stay with the clientsClasses of finding, first audit
- Model output injected into summaries that reach users
- A gap in row-level security
- Server-side request forgery
- Rate limits kept per instance, not per user
No client names, no specifics until fixed and approved.
GovID · Self-initiated, paused
A national-ID brief, argued down from a blockchain to a signed, anchored log.
A red-teamed proof of concept for a digital-identity system. The brief asked for a blockchain; the threat model asked for an append-only log with external anchoring, and that is what I built.
No screenshots: a proof of concept, not a product
Drydock · Own tool, in build
A one-person research desk where no strategy trades until it passes gates set in advance.
Crypto research and paper trading with pre-registered statistical gates, a kill switch and an off-host watchdog. An LLM classifies news; it never places an order.
No profit claims, ever
Everything since 2016, by year
Employer work, small tools and projects still in progress, in one table.