Ontofill
Built at The Agent Arena Hackathon · Sep 26, 2026 · San Francisco, CA
Judges Password: pas89-auar7-gju9z-msyn9 Ontofill is an open-source engine. You give it a one-sentence question, and it runs five phases on its own: it writes a PRD with a testable definition of done, derives an ontology, discovers public sources, plans each source, and sends sandboxed browser and code agents to fill every field. People only approve or deny the PRD, the factors, the ontology and unknown sources. Each decision is bound to the sha256 of the exact file reviewed. Every value keeps its receipt: the source URL, the capture, the locator and the quote. How it's built (Blast Radius Zero): two Vultr VX1 VMs, a control plane and a sandbox host. Every agent LLM call goes through Vultr Serverless Inference, via our gateway, which holds the only key. Each job runs in a gVisor (runsc) cell with memory, CPU, process and time caps. The cell holds zero secrets, can't reach the metadata IP or the mesh, and is destroyed afterwards. Six proof checks are recorded per job. Containment: an extractor that runs rm -rf / and then loops forever is killed by its limits with the host untouched, and a prompt-injection page is quarantined. It covers Pattern B (browser use with Vultr vision verification and an approve-before-submit gate) and Pattern A (self-healing extractors promoted to macros). NetBird: zero inbound ports on both VMs, SSH included. URLs are served through the NetBird reverse proxy, with a password for viewers and SSO for approvers. The VMs talk over WireGuard, one way, and each browser session's live-view URL dies with the session. Two real cases on the same engine: 1. Proveedor Abierto: who receives public contracts in Mexico, and are they legitimate companies? The engine found the federal procurement portal and it renders in the sandbox, but its contract files are served only through a POST with an anti-bot token, so the read-only engine stopped and flagged it by design. There's no engine gold yet; work continues after submission, live on the judges console. A labelled, harness-assisted dataset (394 suppliers, 0 of 7,761 values without evidence) shows the product. 2. San Francisco Public Library branches (Wi-Fi and opening hours): approved through the ontology and now discovering sources. It exposed seven engine defects; six are fixed. Repos: github.com/ricalanis/ontofill (engine) and github.com/ricalanis/proveedor-abierto (case + app). Everything was built during the event, and the commit histories are public.