I'm Nur Amirah Mohd Kamil (Mia), an AI Systems Engineer and the Founder & CEO of MI4 Inc., based in Kuala Lumpur, Malaysia. I work where applied AI meets systems thinking: designing agent architectures, governance layers, retrieval systems, and human-centered products that are stable, explainable, and genuinely useful.
My projects move between enterprise AI reliability and public-good technology, with one constant: intelligence should be engineered as a dependable system—not added as a decorative feature.
“Code is how stardust remembers it was once a star.”
- 🪞 Building Trueframe, a private, on-device studio that makes AI imagery feel more photographic.
- 🛰️ Developing AI-OS and the AI-OS Framework for measurable deployment stability and multi-agent governance.
- 🇲🇾 Exploring AI for public access through BantuRakyat-AI and simulation-led automation for Malaysian SMEs.
| Project | Domain | What it does |
|---|---|---|
| Trueframe · Live | CREATIVE AI |
Six browser-native tools for skin repair, upscaling, film grading, prompt design, and batch processing—with no uploads |
| AI-OS | AI GOVERNANCE |
Introduces ADSI, a composite stability model for monitoring and governing production AI deployments |
| AI-OS Framework | MULTI-AGENT |
Converts operational telemetry into stability scores, risk tiers, anomaly signals, and governance decisions |
| CCAM for Malaysian SMEs | DIGITAL TWIN |
Validates automation workflows inside a deterministic simulation before they reach real operations |
| BantuRakyat-AI · Live | AI FOR GOOD |
Helps Malaysians discover government assistance programs and understand their eligibility |
| ReadyTensor RAG Assistant | RAG |
A FastAPI, LangChain, and Streamlit assistant for context-aware publication discovery |
Everything I ship leaves the notebook in an image — eight Dockerfiles in public repos, thirteen across all my repositories. Each one solves a packaging problem rather than copying a template.
| Repo | The problem it solves |
|---|---|
| ai-os | Two-stage Python build; the runtime layer carries installed packages and source only. Restart, recreate-on-same-volume and backup/restore run in CI — see the lifecycle report |
| Fieldwork | Node build → nginx runtime; the HEALTHCHECK probes a real /healthz location in the nginx config instead of assuming one. Debian build stage because Rollup's native parser ships per-libc |
| miva-studio | CUDA 11.8 / cuDNN runtime for GPU inference, plus a slim preview image for hosts without a GPU |
| midnight-counter | Compose devnet with a deliberately pinned indexer, a proof-server pinned above the arm64 hang, and a bash /dev/tcp healthcheck so --wait means ready |
| CCAM for Malaysian SMEs | uv-based Python build with a .dockerignore that keeps tests and helm charts out of the image |
| ReadyTensor RAG Assistant | Streamlit-first image sized for a Render deploy |
Multi-stage builds by default · healthchecks that can fail a deploy · image tags pinned with the reason written next to them · container lifecycle tested instead of assumed.
Python · TypeScript · JavaScript · FastAPI · React · LangChain · Streamlit · Docker · Docker Compose · GitHub Actions · Google Cloud
Multi-agent architectures · RAG pipelines · AI governance · Digital twins · Containerization · Evaluation · Observability · Privacy-first applications
- Reliability is a product feature. AI systems need measurable behavior, recovery paths, and operational guardrails.
- Validate before reality pays the price. Simulation and deterministic checks belong before deployment.
- Privacy should be architectural. When a task can run locally, user data should stay local.
- Useful beats impressive. The best system solves a real problem clearly and survives contact with its users.
I write up the traps, not the highlight reel — container failures included.
- dev.to/miii — build notes and what broke on the way
- aimirah.com — the longer learning log
I'm open to conversations about dependable AI systems, agent governance, applied RAG, container packaging, creative technology, and mission-driven products.