KAYA Architect

The Autonomous Pre-Construction Analyst.

Comment

Inspiration

When you look at why construction projects get delayed, everyone blames the supply chain, a truck broke down, or a storm hit a port. But when we dug into the data, we found something much worse.

According to the 14th Annual Arcadis Global Construction Disputes Report (2024), the #1 cause of multi-million-dollar construction disputes isn't logistics, it's human error in the contract and procurement phase.

Furthermore, Autodesk and FMI Corporation estimate that this specific type of "bad data" costs the global construction industry $1.85 trillion annually.

We realized that project managers are being asked to manually cross-reference:

  • 50-page, chaotically formatted vendor bids
  • 500-page Master Engineering Specifications

This process is a recipe for human fatigue.

If someone misses a single line item, such as a "specialized thermal coating on a generator", the wrong component gets ordered, causing expensive delays and cost overruns.

Our goal was to build a system that acts as a 24/7 AI procurement analyst, identifying critical scope gaps before contracts are ever signed.

What It Does

KAYA Architect is an autonomous pre-construction analyst powered by Agentic RAG. It executes three core workflows:

1. The Bid Leveler

  • Ingests messy, unstructured vendor bids
    • PDFs
    • Excel files
    • Emails
  • Normalizes them into a clean, side-by-side comparison matrix.
  • Makes it immediately clear:
    • Which vendor included what
    • Which vendor excluded critical scope
    • Pricing differences across vendors

2. The Scope Auditor

  • Cross-references normalized vendor bids against the Master Engineering Specification.
  • Detects missing components automatically.
  • Flags scope anomalies.
  • Autonomously drafts an RFI (Request for Information) email whenever a required specification is missing.

3. The Risk Forecaster

  • Calculates a probabilistic vendor risk score.
  • Evaluates whether a subcontractor poses financial or execution risk.
  • Helps General Contractors avoid signing multi-million-dollar purchase orders with unstable vendors.

How We Built It

We built a highly asynchronous backend using:

  • Python
  • FastAPI

This backend handles computationally intensive ML workloads while communicating with a Neo-Minimalist Next.js frontend through WebSockets, enabling real-time anomaly alerts.

Privacy-First Local AI Stack

Because construction contracts and engineering documents are protected by strict NDAs, we couldn't rely on public LLM APIs.

Instead, we designed an entirely local AI architecture.

Document Intelligence

  • LlamaParse
  • Extracts complex engineering PDFs
  • Preserves nested tables
  • Maintains document structure for downstream reasoning

Memory & Orchestration

  • ChromaDB as the local vector database
  • LangGraph for orchestrating multi-step reasoning workflows

Local Reasoning

Inference is performed entirely using local open-source models via Ollama, including:

  • Llama 3
  • Mistral

This guarantees:

  • Zero proprietary data leakage
  • Complete offline document reasoning
  • Enterprise-grade privacy

Challenges We Ran Into

Parsing Engineering Documents

Our biggest obstacle was document parsing.

Traditional OCR systems completely destroyed the formatting of nested engineering tables, turning carefully structured bids into unusable text.

We ultimately pivoted to advanced document intelligence pipelines that preserved:

  • Table hierarchy
  • Column alignment
  • Engineering numbering
  • Pricing grids

Without this structure, downstream AI reasoning simply wasn't reliable.

Hallucinations in Large Specifications

Standard RAG pipelines struggled with 500-page engineering specifications.

Chunking documents by word count caused the model to lose important engineering context between sections.

To solve this, we engineered a Hierarchical Chunking Pipeline.

Instead of arbitrary token chunks, documents are split using engineering section headers such as:

Section 260000 – Electrical

This ensures the model always understands the engineering context associated with the information it's analyzing, dramatically reducing hallucinations.

Accomplishments That We're Proud Of

One of our biggest achievements was successfully building an autonomous Agentic RAG workflow with minimal hallucinations.

Watching the system:

  1. Ingest a chaotic Excel spreadsheet
  2. Convert it into a strict JSON schema
  3. Query the vector database
  4. Detect that a vendor omitted "seismic bracing"
  5. Automatically draft a professional RFI email requesting clarification

felt genuinely magical.

We successfully bridged the gap between:

  • Messy, unstructured construction documents
  • Structured engineering specifications
  • Reliable AI-driven procurement analysis

What We Learned

Building enterprise AI taught us an important lesson:

The language model is only 10% of the problem. The remaining 90% lies in document parsing, chunking strategy, and data engineering.

We also discovered that the construction industry has an enormous demand for:

  • Private AI
  • Local deployment
  • Secure document processing

Enterprise customers managing billion-dollar projects cannot simply upload proprietary blueprints to public APIs.

Privacy must be built into the architecture from day one.

What's Next for KAYA Architect

Today, KAYA Architect protects construction projects during the Pre-Construction phase by ensuring contracts are accurate before they're signed.

Our next milestone is integrating Architect with KAYA AI's existing post-construction tracking ecosystem (Jarvis and Amber).

The vision is simple:

  • KAYA Architect ensures the correct materials are specified and ordered.
  • Jarvis and Amber track procurement, delivery, and execution throughout construction.

Together, they will create the industry's first truly end-to-end Operating System for Global Construction.

Built With

Share this project:

Updates

Submission history