munim.io · oss.munim.io · LinkedIn
🌟 About Me
I design and ship production AI systems, and I own them end-to-end: architecture, evals, deployment, and the developer tooling around them. My work centres on agentic systems (multi-agent orchestration, tool use, MCP), retrieval you can trust (governed, cited, permission-aware), and real-time voice and conversational agents. I also care a lot about low-resource NLP, because Afaan Oromoo and Amharic deserve the same tooling English gets.
Over three years of delivery for clients in the US, Oman, Japan, and Ethiopia.
🛠 Open Source (oss.munim.io)
Small, sharp tools for building with LLMs. Typed, tested, offline-deterministic, CI-native.
| Tool | What it does |
|---|---|
| codescope | Local-first codebase knowledge-graph MCP server. Symbol graph over 21 languages so agents stop burning tokens re-scanning files. |
| toolsift | MCP proxy that ends tool-definition bloat. Retrieve the few tools you need, not all 50. |
| passmuster | Verify-and-retry loop for model output: generate, check (schema / predicate / judge), retry with feedback. |
| agentvu | Watch your agent think. Record a run, replay it as a live terminal dashboard of tool calls, tokens, and cost. |
| promptsize · tokcost | Context budgets that fail CI, and an offline token + cost counter. |
| oidcdoctor · envvet · gitbroom | Prove OIDC login works, validate .env against its contract, sweep stale branches. |
| heyllm · readmecast · shiptag | LLMs from your terminal, Markdown to animated terminal demos, shareable repo cards. |
Low-resource NLP and research:
- Sentinel-SLM (multilingual moderation and jailbreak guardrail, 1.6M-sample dataset, 20+ languages)
- SagalNet (Afaan Oromoo speech to text)
- Afaan Oromoo news classification
- ArtaiSpellChecker (Amharic, context-aware)
- judgekit (refuse to ship miscalibrated LLM judges)
Also:
- Orbit, my unified calendar, email, todo, finance, and AI hub
- SharAI, source-grounded answers on Islamic legislation
💬 Let's Talk About
- Agentic systems: orchestration, tool use, MCP servers, failure modes, and the boring reliability work that makes them shippable.
- Retrieval architecture: hybrid search, chunk enrichment, provenance, and why the retrieval layer should be your security boundary.
- Evals: golden sets, judge calibration, and gating every change on numbers instead of vibes.
- Voice and conversational AI: latency, interruption handling, persona design, scoring.
- Low-resource NLP: making all of the above work for languages the frontier labs skip.
⚙️ Stack
Python · TypeScript · FastAPI · Postgres + pgvector · Celery / Redis · MCP SDK · LangChain / LangGraph · PyTorch · Docker · Kubernetes · AWS · Azure AI
✨ Fun Fact
When I'm not architecting AI systems, I'm building tools nobody asked for but everybody needs, or arguing that transformers still owe low-resource languages a debt. Proud member of the Survey Corps. 🎯🛡️



