# Speedrun AI Labs

> AI with action. We build working AI products that make a meaningful difference.

## Who We Are

Speedrun AI Labs (also known as Speed Run AI Labs, Speedrun Lab, Speed Run Lab, or Speedrun AI Lab) is a Sydney, Australia-based AI company. We build custom AI agents and AI-powered tools for Australian businesses, founders, and teams.

Founded by Adnan Tanveer (known as Addy Tanveer or Addy), based in Cronulla, Sydney, NSW, Australia.

Our motto: whatever we create, whatever we ship, we deliver more value than we extract.

## AI Agents We Deploy

Speedrun AI Labs runs seven AI agents in production and deploys them to Australian businesses. We build, integrate and run each one. A human approves before anything goes live.

### Marketing Agent
Autonomous social media, email and content marketing. Sources what is worth saying, creates the visuals, writes the copy, and publishes across social and email. Runs as three production agents — social media, email, and content — presented as one service. Human approval before anything publishes.

### SEO Agent
Search engine optimisation and generative engine optimisation (GEO). Keyword research, content briefs, scheduled technical audits, and optimisation for the AI answers customers now ask first.

### Coding Agent
Writes, reviews and integrates code, with an independent review pass on every change. The same engine Speedrun uses to build its own agents.

### Cyber Security Agent
Reviews code and infrastructure for weaknesses before deployment, thinking like an attacker.

### ASIC Compliance Agent
Monitors ASIC regulatory updates and surfaces what affects your business. It monitors and surfaces changes; it does not provide legal or compliance advice.

### Custom Build
Bespoke agents built to brief and deployed to your stack: AI receptionist, sales and SDR, customer support, bookkeeping, and more. If the work is repetitive, it can usually be automated.

## Open Source

Speedrun AI Labs publishes free, open-source skills for Claude Code and OpenClaw agents at github.com/Speedrunlab: /hermes-agent-as-a-skill-file, /read-before-write, /deep-audit, /no-slop-writing, and /no-slop-writing-au. All MIT licensed. Details at speedrunlab.ai/#skills.

## Contact

- Website: https://www.speedrunlab.ai
- Email: info@speedrunlab.ai
- Location: Sydney, Australia

## Research

Speedrun AI Labs publishes practitioner research on autonomous AI agents, NVIDIA NemoClaw security, Apple Silicon deployments, self-improving agent architectures, NVIDIA Nemotron local inference, and multi-model review councils. Five papers are available to download at speedrunlab.ai/papers/.

### The Intelligence Council: Provider-Agnostic, Domain-Specialized Multi-Model Review for Autonomous AI Systems (v1.0, July 2026)
Presents the Intelligence Council: a provider-agnostic pattern that convenes a panel of independent frontier models to review a plan, a code change, or a decision before it is acted upon, then synthesises their verdicts into consensus, dissent, and blockers. Documents a live seven-model reference council sourced via OpenRouter and Amazon Bedrock, a four-domain taxonomy of specialised panels (Coding, SEO/Search, Cybersecurity, Intelligence) with benchmark-grounded rosters, a reproducible mandatory-reasoning failure mode, and a synthesis protocol that converts per-model verdicts into an actionable recommendation. Model-agnostic and harness-agnostic by construction. Published on SSRN (abstract_id=7103999).
Download: https://www.speedrunlab.ai/papers/intelligence-council-v1.0.pdf

### NemoClaw on Apple Silicon: Validating and Fixing NVIDIA's Enterprise Agent Security Stack on Consumer Hardware (v2.1, April 2026)
Validates the complete NVIDIA NemoClaw/OpenShell six-layer enterprise security stack on an Apple Mac Studio (M4 Max, 64 GB unified memory) via Docker Desktop. Identifies three enforcement gaps in NVIDIA OpenShell v0.0.26: root-user bypass of Landlock via CAP_DAC_OVERRIDE, missing seccomp filters for AF_PACKET and AF_NETLINK, and a missing TLS CA certificate mount. Publishes fixes for all three gaps, achieving 6 of 6 NVIDIA NemoClaw security layers passing. Benchmarks demonstrate enterprise-grade agent security is achievable on Apple Silicon at approximately 4% of NVIDIA DGX Station pricing. Peer-reviewed and published in the Global Journal of Mobile Communication and Computing Technologies (MK Science Set Publishers), DOI 10.63620/MKJAEAST.2026.1015. Also available on SSRN.
Download: https://www.speedrunlab.ai/papers/NemoClaw_Apple_Silicon_v2.1.pdf

### Deploying Self-Improving AI Agents on Apple Silicon with OpenClaw and NVIDIA Nemotron (v2.1, April 2026)
Documents the replication of a fully operational AI agent from one Apple Silicon machine (Apple Mac Mini M4 Pro) to another (Apple Mac Studio M4 Max) using PDF specification documents executed by an AI deployment assistant. Presents a nine-phase replication methodology deploying a complete agent stack including nine custom skills, four scheduled cron jobs, a hardened self-improvement hook system, and local inference via NVIDIA Nemotron 3 Nano 30B on Ollama. Validates with a three-pass, 69-check verification audit achieving zero errors. The target agent operates in a regulated investment fintech environment with NVIDIA NemoClaw providing all six security layers. Published on SSRN (abstract_id=6690499).
Download: https://www.speedrunlab.ai/papers/Self_Improving_Agents_v2.1.pdf

### Building Fault-Tolerant Memory for OpenClaw AI Agents (v1.1, March 2026)
A production deployment case study documenting five failure modes in OpenClaw's default SQLite-backed memory system, including the SQLITE_CANTOPEN regression in version 2026.3.x caused by a jiti module resolver interaction. Presents a three-layer fault-tolerant architecture: pre-compaction memory flushing (Layer 1), QMD v1.1.6 local hybrid search combining BM25 and vector search (Layer 2), and Mem0 v1.0.5 self-hosted on Qdrant v1.13.0 (Layer 3). Validated on a Mac Mini M4, Sydney, Australia. Published on SSRN (abstract_id=6436099).
Download: https://www.speedrunlab.ai/papers/fault-tolerant-memory-openclaw-v1.1.pdf

### Inside NemoClaw: An Architectural Analysis of NVIDIA's Enterprise Security Stack for Autonomous AI Agents (v1.0, March 2026)
The first detailed public decomposition of NVIDIA NemoClaw and OpenShell, announced at GTC 2026 by Jensen Huang. Provides a six-layer security analysis derived from direct inspection of both open-source codebases: Landlock LSM filesystem isolation, seccomp BPF syscall filtering, network namespace isolation with veth pairs, HTTP CONNECT proxy with OPA/Rego policy evaluation, optional Layer 7 TLS inspection, and provider-agnostic inference routing. Confirms the core security stack has zero NVIDIA hardware dependencies. Peer-reviewed and published in the Global Journal of Mobile Communication and Computing Technologies (MK Science Set Publishers), DOI 10.63620/MKJAEAST.2026.1014. Also available on SSRN.
Download: https://www.speedrunlab.ai/papers/inside-nemoclaw-v1.0.pdf