CURRENT AI WORKFLOW — WHAT IS LIVE TODAY
NODE 01
CAD INGEST
Filename keyword
parsing (current)
→
NODE 02
MAT_PARSE
AI keyword matching
+ manual override
✓ LIVE
→
NODE 04
AI_COMPLY
Rules-based compliance
routing per vertical
✓ LIVE
→
NODE 05
QUOTE_GEN
AI-optimized pricing
SEC-C engine
✓ LIVE
→
NODE 09
PPA_AUDIT
AI analysis block
in certificate
✓ LIVE
DESIGNED AI WORKFLOW — FULL LLM INTEGRATION (H.22)
INPUT
CAD FILE
STEP/IGES/STL
uploaded by user
→
H.22-A
GEOMETRY LLM
LLM interprets
CAD geometry,
suggests material
→
H.22-B
COMPLIANCE LLM
Checks regulatory
requirements,
flags risks
→
H.22-C
VENDOR MATCH AI
Scores vendors
by capability +
compliance fit
→
H.22-D
PRICE OPTIMIZE
AI negotiates
optimal tier +
delivery mode
→
H.22-E
AUDIT AI
AI generates
compliance summary
for certificate
→
H.23
VECTOR DB
Stores embeddings
for vendor/part
similarity search
LLM MODEL SELECTION
- Current: Claude Sonnet 4.6 (Anthropic) — used for platform development and analysis
- Proof of concept: AI analysis block in Authorization Certificate (SEC-C Optimized label)
- Designed: Claude API integrated into quote.js for real material analysis
- Designed: GPT-4o or Gemini as fallback/comparison model
- Designed: Fine-tuned model on ITAR/AS9100 compliance corpus
VECTOR DB / ML STORAGE (H.23)
- Current: Supabase PostgreSQL — compliance_logs (structured data only)
- Designed: pgvector extension on Supabase for embedding storage
- Designed: Vendor capability embeddings (certification + location + capacity)
- Designed: Part/material embeddings for similarity matching
- Designed: Semantic search across historical transactions
AI SAFETY + AUDITABILITY (H.24)
- Current: Every AI routing decision logged to compliance_logs with auth_id
- Current: Immutable hash chain links every node transition (Claim 18)
- Current: Human override always available (material override control)
- Designed: AI decision explainability layer — why a vendor was matched
- Designed: Confidence scores on material detection and compliance routing
- Designed: Red-flag escalation for ITAR/FDA edge cases — human review required
CURRENT AI PROMPT ARCHITECTURE
- SEC-C AI Analysis block: model, materials, tier, region, auth ID
- Material keyword matching: profile-based assembly classification
- Compliance routing: vertical persona → regulatory framework selection
- Designed: Structured prompts for geometry interpretation
- Designed: Chain-of-thought compliance reasoning per regulation
EXAMPLE DESIGNED LLM PROMPT — NODE 02 MAT_PARSE
SYSTEM: You are a manufacturing compliance AI for ODPS™. Given a CAD file description, identify the most likely material composition for each body/assembly, considering the active regulatory vertical.
USER: File: bracket_assembly.step | Vertical: Aerospace | Detected bodies: 3
Body 1: primary structural, ~45g estimated mass, angular geometry
Body 2: fastener pattern, small cylindrical features
Body 3: surface coating layer
ASSISTANT OUTPUT (structured JSON):
{
"bodies": [
{"id": 1, "material": "Aluminum 7075-T6", "confidence": 0.87, "compliance": "AS9100/NADCAP"},
{"id": 2, "material": "Titanium Grade 5 — ITAR", "confidence": 0.72, "itar_flag": true},
{"id": 3, "material": "Anodize Type III", "confidence": 0.91, "process": "Surface Treatment"}
],
"override_recommended": false,
"itar_review_required": true
}