I build execution infrastructure for AI agents — the layer between a model and the systems it is supposed to operate.

Software engineering student at Telkom University. Available for freelance work in automation, MCP tooling, and browser automation.

zaatool — Notebook-first workflow orchestration for humans and AI agents. Workflows are ordered cells sharing one scope, not node graphs. One notebook definition becomes three things at once: a UI page, an HTTP endpoint, and a tool an agent can call. Exposes 13 MCP tools, with etag-guarded cell edits so two writers cannot silently overwrite each other. React 19 Express 5 PostgreSQL MCP

rewind-bench — A benchmark separating repairs that passed from repairs that held up. Bugs are injected by AST mutation; every accepted repair is re-run against a hidden second dataset. Repairs that only match the observed data are flagged as lucky passes. It found that 7% of accepted repairs from the best-equipped agent were fake — they matched recorded output and broke on new inputs. 297 recorded episodes; every number regenerates from committed data with no inference spend. TypeScript AST (Acorn) Vitest MCP Read the report →

nb-steprunner — Any web page, as a Jupyter-style notebook. Write and run JavaScript cells against the live DOM with shared state. Checkpoints survive page reloads, so a long automation resumes from its last successful step instead of starting over. Reaches into cross-origin iframes, and runs on sites whose Content-Security-Policy blocks eval. Preact Vite CodeMirror 6

I document what my tools cannot do as carefully as what they can. On rewind-bench my preferred approach won 30/30 against 28/30 — while burning 3.7x more tokens, so I reported both numbers. I would rather you plan around a real limit than discover it in production.

Certified: IBM RAG and Agentic AI · Google IT Automation with Python (2026) Reach me: LinkedIn · afriza.gilleon@gmail.com