The builder behind hamz.ai

Hamza Malik

hamz.ai is the AI implementation studio. This page is the person behind it: I build production Claude systems, and I hold my own work to the same standard the studio's lab holds client work to. Verified numbers, open repos, and what I'm actually building this month.

Reliable AI Skills → The Lab → GitHub LinkedIn
01Now

What I'm building this month

Updated monthly. The date stamp is the point.

Changelog2026-07
+Reliable AI Skills is public: 17 open-source Claude Code skills, MIT licensed, shipped with the harness and pre-registration that graded them.
+In San Francisco through July 22, then back August 6-17. Say hi if that's you.
~MIT Sloan MBA, graduating May 2027. Research and Teaching Assistant with Prof. Kate Kellogg on a new Fall 2026 course, Leading Agentic AI Transformation in Organizations. I'm working on three sessions: interviews with domain and technical experts, building the cases and exhibits, and defining the class workflow and learning outcomes.
~Available full-time June 2027.
02Proof

Only numbers I can point at

No ROI slides. Each of these traces to a repo, a test run, or a dated measurement.

26/50 → 46/50task must-hits, order-blinded skills A/B across 16 tasks (2026-07-02)
689automated test files behind Lyna Health, measured 2026-07-09
32tracked AI call expressions across 20 modules
17open-source skills, consolidated from 26
$965.22total cost of the effort-lattice study · full write-up →
2production repos Claude Code runs in

689 and 32 are from Lyna Health, one of the two production repos below. 17 skills: 26 repo-specific skills consolidated into 15 generalized skills (19 merged down to 9, 6 carried 1:1, 1 dropped as product-specific, 2 largely absorbed into merges), plus 2 promoted from a separate personal library. Only 5 of the 17 carry measured behavioral lineage from the original A/B; the rest are unmeasured or net-new.

03Built

Things I build

Open them, look around, feel free to reach out.

github →

Reliable AI Skills

17 Claude Code skills, MIT licensed, shipped with the eval harness that graded them and the pre-registration frozen before the runs.

The repo → Browse the skills →
the lab →

The Lab

The lab of hamz.ai: where I test how AI systems actually behave, and the standard every client deliverable passes before it ships.

Open the lab → The effort lattice →
field guide →

Designing agentic systems

One agent or many, where a step should stay plain code instead of a model, and how context allocation, model routing, and orchestration map together.

Read it → The 3D version →
github →

Clara

Reads dense official mail and explains it back in plain language. Built at an MIT hackathon run by the Claude Builders Club, sponsored by Anthropic and Jane Street, build weekend May 2026.

The repo →
production →

Lyna Health

A production healthcare product. The evidence harness on the Assurance Chain runs against it for real: 689 automated test files, 32 tracked AI call expressions across 20 modules, every model swap decided by a shadow-mode comparison before it ships.

The harness →
production →

Cubicle

A second production app I build for. The tiered-escalation climb on the Assurance Chain, and one of the adversarial-review receipts, are both real rounds from here.

The climb →
client work

Capital markets advisory platform

An advisor cockpit and client portal I built for a capital markets advisory firm in NYC (client anonymized), with an overnight AI pipeline that turns meeting and session material into structured notes, carried-forward tasks, and drafted client messages, all behind a manual approval gate. Nothing sends without the advisor's sign-off, and nothing gets invented: every extracted claim carries a verified source quote.