
FactoryForge is a multi-agent software factory designed to automate the software-building process through coordinated AI agents. Instead of asking one agent to handle an entire project, FactoryForge divides the work into explicit, independent work packages with clear ownership, dependencies, deliverables, acceptance criteria, and non-goals. A Planner analyzes the specification and decomposes it into executable packages. Implementers work only within their assigned scope, while Test Authors independently create verification checks. Reviewers inspect the resulting changes and evidence before work is considered complete. The factory is built around explicit agent-to-agent handoffs rather than relying on implicit context. Shared contracts and disjoint file ownership reduce conflicts between parallel agents, while acceptance commands provide concrete evidence that a package actually works. Our Dark Factory demonstration uses BAND Desktop to coordinate the agent team through rooms, tasks, artifacts, and structured communication. The agents collaboratively build and verify a software service while respecting the constraints defined by the specification. The goal of FactoryForge is to demonstrate a reusable software-production system where planning, implementation, testing, review, and verification become coordinated stages of an autonomous development workflow rather than a single AI coding session.
6 Oct 2026

KitchenGuard is a hands-free, voice-native food-safety inspection system designed for commercial kitchens. Instead of forcing workers to stop, type, and complete long inspection forms, KitchenGuard lets them walk through the kitchen and speak naturally. AssemblyAI powers the real-time voice interaction, while KitchenGuard converts spoken observations into structured inspection data. The system uses a deterministic inspection engine to evaluate configurable rules, detect missing information, identify issues, and prevent an inspection from being finalized before required checks are completed. KitchenGuard also includes persistent kitchen memory. For example, if a worker says, “There is a blue cleaning bucket under the refrigerator,” that spatial relationship is stored as structured memory. Later, the worker can ask where the bucket is, update its location, or retrieve its previous location. Corrections are tracked rather than silently overwritten, creating an audit history of observations, changes, issues, and resolutions. The result is a voice-first inspection workflow that combines real-time speech, structured actions, deterministic verification, persistent memory, and report generation.
30 Sep 2026