HoldPoint

HoldPoint detects hidden construction revision conflicts, stops affected work, generates corrective control documents, and re-verifies the project before work resumes.

Comment 1

Inspiration

Construction documents are not dangerous independently.

The failure happens in their relationships.

A drawing can still look valid even after a newer RFI has already changed one scoped requirement. That is where costly rework, delays, and field confusion begin. We wanted to build a system that catches that hidden conflict before it becomes a physical mistake on site.

What it does

HoldPoint analyzes new project instructions such as RFIs, extracts structured claims with independent AI models, compares them against the active control set, determines precedence with a deterministic engine, and decides whether work can proceed, must be held, or needs review.

When a conflict is confirmed, HoldPoint:

  • identifies the authoritative requirement
  • shows the affected work packages and downstream blast radius
  • changes the project state from PROCEED to HOLD
  • generates an AI Resolution Pack
  • creates a dynamic Control Pack PDF
  • supports simulated engineer approval and deterministic re-verification until the project returns to PROCEED

How we built it

We built HoldPoint as a Next.js and TypeScript application with a multi-model inference layer and a deterministic decision engine.

AI layer

We use:

  • DeepSeek V4 Flash for primary extraction
  • Qwen 3.8 27B for independent verification
  • GLM 5.2 for escalation when the first two materially disagree

Deterministic layer

Our TypeScript engine handles:

  • claim normalization
  • scope matching
  • revision precedence
  • explicit supersession
  • contradiction detection
  • dependency propagation
  • final PROCEED / HOLD / NEEDS REVIEW decisions

Actionable output

After a confirmed conflict, HoldPoint generates:

  • a Drawing Revision Notice
  • an RFI Resolution Response
  • a Field Action Checklist
  • a downloadable Control Pack PDF

These documents are clearly labeled AI DRAFT / ENGINEER APPROVAL REQUIRED. AI can draft corrective material, but it cannot approve its own output or release the job back to construction.

Why this is not an LLM wrapper

HoldPoint does not let a model decide project authority.

AI is used to interpret raw construction language and extract evidence-backed structured claims.

The deterministic engine decides:

  • which document is authoritative
  • whether work must stop
  • which downstream tasks become blocked
  • whether the project can return to PROCEED

That separation was one of the most important design decisions in the project.

Core workflow

The main demo flow is:

Drawing A12 Rev 3 250 mm → Raw RFI-047 → DeepSeek V4 Flash + Qwen 3.8 → 2/2 extraction consensus → Deterministic precedence resolution → 300 mm becomes authoritative → Dependency propagation → HOLD POUR → AI Resolution Pack → Control Pack PDF → Simulated engineer approval → Deterministic re-verification → PROCEED

Challenges we ran into

The hardest challenge was building something that felt like a real operational system instead of a one-click AI demo.

We had to:

  • separate language understanding from decision authority
  • validate model output safely with schemas
  • handle disagreement between models
  • create a believable approval boundary
  • generate useful corrective artifacts instead of generic AI text

We also had to make sure secrets stayed server-side and that the demo still remained usable in fallback mode.

What we learned

We learned that strong AI products often come from combining probabilistic understanding with deterministic operational logic.

The useful part was not just extracting information. The useful part was turning that information into a decision, a blast radius, a corrective document set, and a verified release path back to work.

What's next

Next, we would like to connect HoldPoint to real document-control systems, richer construction datasets, and more complex approval workflows.

We would also like to support broader coordination scenarios beyond the current synthetic demo dataset.

Built for Impact Forge Summer 2026

HoldPoint was built as a decision-support prototype for Impact Forge Summer 2026. It is a hackathon MVP and a decision-support system, not a substitute for professional engineering review.

Built With

Share this project:

Updates

Submission history