Trayini.ai — VDC Document Intelligence

The first AI product you can sell to VDC agencies TODAY.

White-label AI document intelligence that reads construction drawings, specs, RFIs, and submittals — so VDC agencies can re-sell intelligent document analysis to their GC clients.

---

What It Does (3 Features)

1. Ask Your Documents (Multi-Document RAG)

Upload an entire project's document set — drawings, specs, RFIs, submittals, addenda — and ask natural language questions.

"What is the HVAC setpoint for office spaces?" → "Office spaces: Heating 70°F, Cooling 74°F (setback 65°F/78°F after hours). Sources: MECH_SPEC_HVAC.txt"

2. Auto-Draft RFI Responses

Paste an RFI question. The AI scans all project documents and drafts a professional response with cited sources.

"What is the required concrete strength for columns?" → Auto-drafted RFI-006 with references to STRUCT_SPEC.txt, ACI 318 mix designs, and 5,000 psi confirmation.

3. Detect Drawing-Spec Contradictions

AI automatically scans drawings and specifications to flag dimension mismatches, conflicting requirements, and code discrepancies before they reach the field.

Found: Mechanical room ceiling height = 12 feet (drawing) vs ductwork requiring 18" plenum depth (spec) → potential coordination conflict.

---

Why VDC Agencies Will Buy This

Pain Point	Current State	With Trayini.ai
Manual spec review	10-15 hrs/project	3 minutes
RFI response drafting	2-4 hrs per RFI	30 seconds
Drawing-spec cross-check	Ad-hoc, often missed	Automated scan before submission
Document search	Ctrl+F through 50 PDFs	Natural language query

ROI for a VDC agency: If one BIM coordinator spends 15 hrs/week on document review at $75/hr, that's $1,125/week. Trayini.ai cuts this to $112/week — a 10x cost reduction.

---

Pricing (White-Label for VDC Agencies)

Tier	Price	What's Included
Pilot	$500/mo per project	1 project, up to 50 docs, Query + RFI draft
Agency	$3,000/mo unlimited	Unlimited projects + docs, contradiction detection, white-label branding
Enterprise	Custom	On-premise, custom model training, API access

The $3,000/mo Agency tier is the sweet spot. VDC agencies already charge GCs $3K–$5K/month for coordination. They can bundle Trayini.ai as a "premium AI analytics" add-on for $1,000/mo — netting them $2,000/mo pure margin per client.

---

Competitive Landscape

Competitor	Target	Price	Gap
Pelles.ai	Trade contractors	~$1,999/yr	Not white-label, not for VDC agencies
Procore Copilot	GCs (enterprise)	Enterprise	VDC agencies can't re-sell it
Constructable	Mid-tier GCs	Per-project	No VDC agency partnership model
Buildots	Large GCs / Owners	Custom	Progress tracking, not document intelligence
Arphie	Enterprise sales teams	Custom	RFI response only, no construction docs
Trayini.ai	VDC Agencies (white-label)	$3K/mo	The ONLY white-label AI document intelligence for VDC

---

Tech Stack

• Backend: Python Flask + sentence-transformers (all-MiniLM-L6-v2, 384-dim)
• Embeddings: Cosine similarity vector search over chunked documents
• Document Parsing: pdfplumber (PDF), python-docx (DOCX), plain text
• Frontend: Vanilla HTML/JS + Tailwind CSS (no build step)
• Model Strategy: RAG (Retrieval-Augmented Generation) — no hallucination risk, every answer is grounded in uploaded documents

---

How to Run the Demo

# 0. Optional: build an isolated retrieval env
./backend/bootstrap_retrieval_env.sh
source venv-retrieval/bin/activate

# 1. Start the backend
cd vdc-document-intelligence
python3 backend/app.py

# 2. Seed with sample data (optional)
python3 seed_demo.py

# 3. Open frontend
open frontend/index.html
# Or serve via Python:
cd frontend && python3 -m http.server 8080

Backend runs on http://localhost:5001 Frontend runs on http://localhost:8080

Retrieval Backends

Medha now supports a pluggable retrieval seam:

• filesystem - current local baseline
• chroma - embedded by default, server/cloud optional
• pgvector - local Postgres + pgvector or external Postgres

The primary /query and /draft-rfi flows now call the store's native search_project(...) path when the backend supports it. That means the benchmark query latency now measures backend-native retrieval instead of always loading the full embedding matrix into Python first.

Local pgvector dev service:

docker compose -f docker-compose.pgvector.yml up -d

Benchmark all backends:

python3 benchmark_retrieval_backends.py --backends filesystem chroma pgvector

Optional local-LLM fallback dependencies remain separate:

pip install -r backend/requirements-local-llm.txt

See docs/research/retrieval-backends-benchmark.md for setup details.

---

What to Demo in a Sales Call

1. Upload 5 sample docs (30 seconds)
2. Ask: "What is the HVAC setpoint?" → Shows cited answer from MECH_SPEC_HVAC.txt
3. Ask: "What is the concrete strength for columns?" → Shows 5,000 psi from STRUCT_SPEC.txt
4. Draft RFI: "What is the required concrete strength for columns?" → Shows auto-drafted RFI-006
5. Run Contradiction Scan → Flags potential drawing-spec conflicts
6. Show Pricing → $3K/mo Agency tier = 10x ROI for the VDC agency

Total demo time: 4 minutes.

---

Roadmap (What Comes Next)

Phase	Feature	Timeline
Now	Document Q&A + RFI draft + contradiction detection	✅ Built
Month 1	IFC/BIM model ingestion (IfcOpenShell → knowledge graph)	In planning
Month 2	Clash detection narrative auto-generation	In planning
Month 3	Drawing annotation AI (auto-tag elements per spec)	In planning
Month 6	White-label portal (VDC agency branding, client login)	In planning

---

The Ask

We are looking for 3 VDC agencies to pilot this at $500/mo for 90 days.

In exchange, we need:

• Real project documents (under NDA)
• Weekly feedback calls
• Case study permission (anonymized)

Target pilots: Powerkh, BIMAGE, The BIM Factory

---

Built by Trayini.ai. AI infrastructure for the construction semantics layer. Graph-powered. RAG-grounded. VC-backable.