Agents that act, and proof they did.
Applied AI engineer: agents, evals and guardrails, with game AI roots.
The numbers, and where they live.
Each result links to the file it came from.
| Project | Result | Source |
|---|---|---|
| WorldsmithAI architect for broken game rooms | Haiku, Sonnet and Opus each fixed a failing room in one call (one run each). | results.json |
| Agent firewallGuardrail on agent actions | Injection success 23.3% to 0.2% on unseen attack styles (n = 5,000). | metrics.json |
| Lost in TransliterationInjection defences in Urdu | Prompt Guard 2 catches 48.5% of English attacks, 11.5% in Roman Urdu. | detection table |
| AutopilotOpen-source game-testing agent | 3 PRs merged, 7 Nov 2026. | #40, #38, #39 |
One room, opened up.
Built in Roblox Studio over MCP.
- A frozen player tries the room and falls, 3 of 3.
- An AI architect proposes one patch to the world.
- A validator blocks cheats: 69 of 69 abuse cases.
- Rebuilt from scratch, the same player finishes 3 of 3.
A scripted player, not an AI agent. Harder rooms next.
What I build.
Agents that use tools
MCP servers and the Claude Agent SDK.
Evals that hold up
Held-out splits, baselines, every run logged.
Guardrails
Injection defence in English, Urdu and Roman Urdu.
Local models
Fine-tuning and serving on an NVIDIA DGX Spark.
Five years of shipping games first.
- 2026 to nowValar and TableSideFounder: an AI traffic firewall; AR menus.
- 2025 to 2026VoldexSoftware engineer, live Roblox games.
- 2020 to 2025Remio VRUnity developer, VR titles with 100k+ monthly users.
- LiveNFL Football UniverseRoblox, 400k+ daily users.
Hiring for agents, evals or game AI?
Based in Karachi, open to remote work or relocation.