I built OG Foreman so one construction update can do more than sit in a chat

Daries, a computer science student and AI engineer from Accra, Ghana, has developed an artificial intelligence tool called OG Foreman to streamline construction site updates. In the construction industry, progress updates often consist of short messages that require multiple follow-up actions. OG Foreman, an AI construction site coordinator, addresses this issue by connecting site updates to actual tasks, materials, dependencies, and project records.

When a site team reports an update using text, voice, or photos, OG Foreman interprets the information and performs several actions. For example, if a message states that "the electrician did not come," OG Foreman records this absence as a blocker rather than marking electrical work as complete. The tool then updates the cement stock, compares it with the requirements for plastering, prepares a material request for the shortage, and refreshes the Daily Log and activity history. The construction site manager must approve any consequential actions before they are implemented.

OG Foreman's architecture is composed of four key components. Gemini 3.6 Flash reads project sources and interprets authorized site evidence, generating structured information without directly modifying the project database. The Google Agent Development Kit (ADK) coordinates operational workflows, including situations where an action must wait for approval.

Typed tools and application services handle identity, permissions, project references, safety rules, and duplicate requests before making any changes. Project truth and the audit history are stored in Cloud Firestore, while the web app is built using Next.js, React, TypeScript, Firebase Authentication, and Firebase Hosting. The FastAPI API and private worker run on Cloud Run, and Pub/Sub moves durable project events to the worker, while Cloud Storage holds private media.

Lastly, Cloud Logging and Cloud Trace connect the visible project results to the workflow that produced them.

Daries emphasizes the importance of clearly defining what the AI agent can do and specifying when a human must intervene, ensuring routine record-keeping moves automatically while purchases, external commitments, schedule changes, cancellations, and safety-critical decisions remain under human control. The developer is currently working on a broader public beta with more repeatable live evaluations, multilingual site updates, manager-reviewed corrections, and clearer media-retention controls. The source code and architecture documentation are available on GitHub.