AI/ML Engineer — LLM & Agentic AI Systems, security-first
I build production-shaped LLM agent systems and care about the parts that break in production: model-output validation, untrusted-input boundaries, and numerical correctness. Python across the agentic stack — LLM orchestration, MCP tooling, CI/CD on cloud — with a security-research backbone from hardening real open-source codebases.
📍 Indonesia · open to remote AI/ML roles & freelance/contract · ✉ naufalhfzh805@gmail.com
Python · pytest · LLM agents · Multi-agent orchestration · MCP ·
Model-output validation · Prompt-injection defense · Untrusted-input boundaries ·
REST APIs / SDKs · Google Cloud Run · GitHub Actions · GPT-4o / OpenAI ·
TypeScript / Node.js · Rust (Soroban)
🔥 Firebreak — Autonomous Cloud-Ops LLM Agent An autonomous agent that diagnoses Google Cloud Run incidents, executes controlled remediation from a typed action catalog, verifies the result, and rolls back on failure. Safety-first: model-output validation, explicit safety policy, an audit ledger of every action, concurrency control. 529 automated tests incl. chaos testing — zero unauthorized mutations across the suite.
🧮 Numerical-correctness research — tenstorrent/tt-metal
Independently found and root-caused a floating-point cancellation defect in
ttnn.hardsigmoid / ttnn.hardswish — up to 25% relative error on the FP32 path,
reaching MobileNetV3 / EfficientNet-class workloads. Built a dependency-free IEEE-754 model
to prove it; the proposed fix was adopted verbatim into an official maintainer-track PR.
🏅 MCP Sentinel — Top 4 of 296, lablab.ai AI Factory Hackathon (2026) An MCP security proxy that validates tool calls between agents and MCP servers. Placed #4 of 296 shipped projects by community vote at the AI Factory hackathon (3,711 participants, 1,122 teams).
🧠 MCP & LLM tooling — memanto (moorcheh-ai) Guarded MCP batch operations against malformed results; fixed noisy LLM JSON extraction — the untrusted-output boundary work that keeps agent pipelines from silently corrupting.
🛡 Security & correctness — 15 merged PRs across production OSS
- TestSprite CLI — 10 merged PRs; recognized in the Season 3 CLI Improvement Track. Locked credential-profile mutations against cross-process races, tightened Windows credential ACLs, stopped non-idempotent retries, rejected malformed API keys before dispatch, guarded secret-file reads behind typed validation.
- TrusTrove (Soroban / Rust) — hardened a pool-utilization overflow in a smart contract.
- GitHub Actions OIDC research labs:
oidc-caller-lab·oidc-callee-lab
Reproduce first, report second · minimal, reviewable fixes · clear evidence over broad claims.
D4 (Applied Bachelor), Biomedical / Electromedical Engineering · expected 2027 · AI Factory certified (GPT-4o / OpenAI, 2026)