keystone

An AI-native Promise Intelligence Engine mapping construction dependencies onto a dynamic graph to predict broken commitments, quantify risk cascades, and simulate optimal fixes in real time.

Comment

Inspiration

Massive construction projects rarely fail because a single truck is late; they collapse because thousands of invisible, interdependent commitments silently break[cite: 1]. Traditional project management software operates like an autopsy report—it only shows you the damage after the project is already hit with penalties[cite: 1]. We built KEYSTONE to give construction a brain, moving heavy industry from reactive firefighting to proactive control[cite: 1].

What it does

KEYSTONE introduces an AI-native Promise Intelligence Engine that transforms every deadline, material delivery, and site inspection into a living, measurable object mapped onto a dynamic Promise Graph[cite: 1].

When automated signals flag a statistically significant confidence drop—such as a steel delivery slipping from 90% to 62%—KEYSTONE jumps into action[cite: 1]. Our Cascade Engine calculates the precise downstream ripple effects across the dependency graph (like idle labor costs and contract penalties)[cite: 1]. Simultaneously, our Counterfactual Simulator runs a constrained search over discrete recovery actions (expediting, substituting vendors, or re-sequencing crews) to surface the single cheapest fix[cite: 1]. Finally, autonomous agents close the loop by drafting vendor emails and updating schedules[cite: 1].

How we built it

We engineered KEYSTONE using four core pillars mapped to a unified graph architecture:

  • Document Intelligence: Vision-language models that parse POs, submittals, and fabrication schedules to build and update graph nodes automatically[cite: 1].
  • Forecasting Engine: Computes real-time probabilistic estimates over delivery and inspection timelines[cite: 1].
  • Graph Optimization Layer: Extends traditional Critical Path Method math into a dependency DAG to propagate risk edges algorithmically rather than relying on LLM guesses[cite: 1].
  • Agentic Execution Layer: Multi-agent framework capable of executing outbound vendor coordination and schedule syncs[cite: 1].

Challenges we ran into

Generalizing the Critical Path Method math to handle live, probabilistic confidence scores over a complex Directed Acyclic Graph (DAG) was a massive mathematical challenge[cite: 1]. We had to ensure the Cascade Engine could accurately propagate risk edge-by-edge without incurring massive latency, enabling the Counterfactual Simulator to score thousands of permutations in milliseconds[cite: 1].

Accomplishments that we're proud of

We successfully moved AI beyond a simple "chatbot wrapper" into a deeply integrated operational intelligence system[cite: 1]. Our platform handles complex industrial bottlenecks natively, proving that AI can accurately calculate real-world financial risk exposure and automate physical supply chain mitigation strategies[cite: 1].

What we learned

We learned that construction teams don't need another generic alert dashboard or text summarizer[cite: 1]. In heavy industry, predictability and actionable alternatives are the ultimate moat[cite: 1]. True intelligence isn't just predicting a delay—it's executing the optimal detour before the impact hits the ground[cite: 1].

What's next for KEYSTONE - by code for kaya

We plan to scale KEYSTONE from single-site implementations to an enterprise-wide portfolio view, seamlessly linking multiple projects, cross-site crews, and shared vendors without changing the underlying graph schema[cite: 1]. We will also integrate live telematics and predictive weather models directly into our forecasting layer to make promise tracking even more autonomous[cite: 1].

Built With

Share this project:

Updates

Submission history