1
1
Viet Nam
6+ years of experience
Software engineer passionate about developer tools, backend systems, and AI-assisted workflows. Interested in building practical AI experiences that help developers automate repetitive tasks, better understand complex codebases, and ship software more effectively. Currently exploring AI agents, workflow automation, and developer-focused tooling that makes AI a more reliable engineering partner in real-world software development.

ATOMiK is a constitutional control plane for coding agents: it does not just show agents doing work; it shows them being constrained. The demo seeds a subtle bad PR that bypasses a protected-ref policy. A baseline reviewer approves it. ATOMiK then runs three legislative-branch agents, commits structured proposal artifacts on real Git worktrees, uses git merge-tree --write-tree to produce executive candidate state, scores the delta through an ATOMiK C ABI stub, and sends governance metadata to a Lean-backed judiciary. The judiciary vetoes unsafe dispatch, preserves the protected ref, and records the audit trail in JSONL, Git metadata, signed/simulated constitutional epoch tags, and a live dashboard. The model layer targets AMD MI300X-hosted coding-agent inference via ROCm and an OpenAI-compatible vLLM endpoint, with deterministic fallback to keep the governance proof reliable during judging. Business buyers are platform engineering, AppSec, and DevSecOps teams adopting coding agents in repositories where protected branches, audit evidence, and rollback matter. The first product is a GitHub/GitLab app that blocks agent-authored protected-ref or policy-sensitive changes unless they include a signed judiciary ruling artifact. Orchestrators route tasks, AgentOps tools observe traces, and guardrails flag risky text. ATOMiK decides whether an agent-authored repository state transition may advance protected authority. The model can change; the constitution still wins.
10 May 2026