AI Engineer · Full Stack Architect · Founder @ NorahLabs

Building production AI systems from Nairobi, Kenya. Coding-agent harnesses, agentic orchestration, and native desktop applications — designed for developers who want serious AI tooling that runs on open-weight and local models, without cloud lock-in.

The Open-Weight Frontier Harness — the agent workstation built for open-weight frontier labs, where a weaker local model performs like it knows your codebase.

macOS / Linux / Windows · atlarix.dev

Atlarix is built for the open-weight frontier labs — DeepSeek, Qwen, Kimi, MiniMax — not just compatible with them. The bet: durable open-weight labs, not specific model versions. What makes a weaker model punch above its weight isn't a bigger prompt — it's the harness. Enforced tool approvals, an OS-level execution sandbox, and verified edits mean the model's mistakes are caught by the system, not trusted on faith. Any-model BYOK, local-first execution, every destructive action behind your approval.

What makes it different:

• Built for open-weight, not retrofitted — DeepSeek / Qwen / Kimi / MiniMax as first-class, plus broad BYOK (OpenAI, Anthropic, Gemini, OpenRouter, Groq, Together, Mistral, xAI, Hugging Face) and local Ollama / LM Studio.
• Harness-managed agent control — tool approvals, background commands, waits, and sub-agents are enforced by the harness, not the prompt, so a weaker model can't emit a premature completion or mismanage a wait.
• Verified edits — in Build / Debug the agent runs the project's own checks (tsc/eslint/ruff/mypy/pytest) through a sandboxed terminal and can't declare a task done while they fail.
• OS-level execution sandbox — per-OS write-confining command execution (Linux Landlock, Windows AppContainer, macOS Seatbelt), an approval queue with hunk-level diff accept/reject, a danger gate, and committable permission + hook rules at a single execution funnel.
• Fast search, no index — bundled-ripgrep grep and glob over your workspace: no index to build, no background watcher, constant low memory at any repo size.
• Atlarix Core — managed open-weight models with a pay-as-you-go credit wallet; no API key required, full BYOK / local on the free tier.
• Parallel sub-agents — the task tool fans out up to five concurrent read-only scouts per turn, live thinking streamed per agent.
• Token-efficient by design — on-demand tool-driven retrieval and model-triggered compress_context; provider-native tool blocks and prompt caching lift weak / local model performance.
• Real workspace — interactive PTY terminal (persists across restarts), in-app browser, @ file/folder mentions scoped to the current workspace, per-workspace MCP.

Work modes: Explore (read-only) · Plan · Build · Debug · Review (correctness + security)

As an open-source contributor, I've landed quality work into major repos using Atlarix on open-weight models — including Qwen Code and Kilo Code, a directly competing coding agent.

Achievements: Winner — Amazon Nova AI Hackathon (Devpost × AWS, 2026) · GDG Nairobi Agentathon (2026) · Lua × Antler (Nairobi, 2026) · Red Bull Basement Global Innovation Programme

Blueprint: Section-Scoped Structural Graph Retrieval and Post-Turn Compression for Agentic LLM Coding in Multi-Repository Workspaces Amariah Kamau, NorahLabs — 2026

Documents the Blueprint architecture — a section-scoped structural index (Universal Ctags symbol graph + ast-grep edges + SQLite FTS5) that handed the agent structural understanding in ~6,500 tokens instead of a whole-repo dump — alongside post-turn tool-result summarisation and benchmark results from a controlled A/B evaluation on a production multi-repository workspace, including a counterintuitive finding: structural confidence lets the agent explore more, using more total context but producing stronger results.

atlarix-skills — Community Agent Behaviors registry. SKILL.md files teaching agents language patterns and framework conventions across React, Next.js, Python, TypeScript, Go, Rust, Docker, MCP, and more. Apache 2.0.

atlarix-mcps — MCP server registry for Atlarix. One-click connections to Gmail, Calendar, Drive, and more from within the agent environment. Apache 2.0.

atlarix-releases — Desktop app release builds for macOS, Linux, and Windows.

Praxia — AI healthcare assistant with MONAI X-ray analysis (pneumonia, fractures, tumours), multilingual symptom diagnosis, real-time WebSocket chat, and a Docker/Nginx/Celery production stack.

Commit Checker — Open source CLI for Git commit analysis.

</file_text>