What Signal Finds in Engineering Firms
Your specs, submittals, and contracts contain risks your team doesn't have time to catch. Signal finds them automatically, on your hardware, before they become change orders.
What Signal for AEC Is
Local-first AI platform for Architecture, Engineering, and Construction firms. Reads any document a firm produces or receives, specs, submittals, contracts, RFIs, inspection reports, drawings, and turns it into structured, searchable, verified data.
No cloud. No per-seat SaaS. Runs on the firm's own hardware.
Value delivered:
- Recovers $300K-$500K annually per firm (CAD rework reduction alone)
- Replaces 2 days of junior engineer manual spec review with 30-second AI processing
- Captures institutional knowledge when senior engineers retire
- Ensures compliance by cross-referencing jurisdiction-adopted building code editions
- Zero cloud API exposure for firms with NDA-protected data
Competitive Moat
1. Air-Gapped Architecture
The only local-first product. Competitors require cloud. Engineering firms with NDA-protected data will not accept cloud. Non-negotiable for utilities under NERC CIP.
2. Jurisdiction Code Registry
Structured data on which authority having jurisdiction adopted which code edition with amendments. Does not exist elsewhere in queryable form.
3. Domain-Specific AI
Understands ASCE 7, ACI 318, OSHA 29 CFR 1926, and inter-entity relationships. Not generic document AI.
4. Consulting-First Positioning
Assessment engagement removes procurement friction, proves value before software commitment. Engineer-to-engineer, peer not vendor.
5. Founder Domain Expertise
13+ years in geotechnical and civil engineering. Ran a firm. Adopted AI tools and validated them across 200 engineers. Instant credibility with buyers.
Architecture
Three engines orchestrated through a FastAPI server with workflow orchestration:
| Engine | Function | Scale |
|---|---|---|
| Vanta | Parse, classify, enrich, decompose, index | 13 files, ~12,600 LOC |
| Aletheia | Validate, audit, jurisdiction cross-reference, certify | 6 files, ~4,100 LOC |
| Daedalus | Retrieve patterns, generate reports, produce CAD/BIM scripts | 6 files, ~3,200 LOC |
~19,900 lines of production code. 554 tests across 36 files. 70% minimum coverage.
The Pipeline
Six stages. Every document passes through the same pipeline.
PARSE → PRE-FILTER → CLASSIFY → ENRICH → DECOMPOSE → INDEX
| Stage | What Happens |
|---|---|
| Parse | 16+ formats (PDF, DOCX, DXF, CSV, images via OCR). Returns raw text, cleaned text, metadata. |
| Pre-Filter | Structured excerpt of high-value units with authority/risk profiles and standards as confidence signals. |
| Classify | Deterministic-first: filename + regex + decompose signals scored. Confidence above 0.8 skips AI entirely. Falls through to local LLM only when signals are weak. |
| Enrich | Five built-in plugins: standards cross-reference, PII detection, timeline extraction, financial analysis, contract parsing. |
| Decompose | Header-aware markdown chunks or character-based with overlap. Per-unit classification (authority, risk, domain). |
| Index | Embed with local models (768d or 384d) and upsert to vector store. Three collections: documents, geometry, jurisdictions. |
Processing benchmarks: small doc (10 pages) in 2-5s, medium doc (50 pages) in 10-15s, large doc (200+ pages) in 30-60s.
Three Engines
Vanta: Document Intelligence
Text extraction from 16+ formats. Deterministic-first classification with 62 AEC keywords and 60+ document type signatures. Confidence-gated AI that only calls LLM when rule-based confidence is below 0.70, using three-channel parity (regex + structure + AI majority vote).
Five enrichment plugins with circuit breaker (3-failure quarantine). Local embeddings with automatic fallback. Parallel batch processing with content-hash deduplication.
Aletheia: Verification and Audit
Schema validation producing quality scores (0-100) with field-level pass/fail and cross-field invariants. Safety-critical keywords receive 5x penalty weight.
Jurisdiction code registry with 500+ code bodies seeded. Cross-reference engine verifying adopted/superseded/not-found status. Hash-chained audit ledger with tamper-evident certificates at gold/silver/bronze levels.
Daedalus: Retrieval and Generation
Vector search with jurisdiction, discipline, and risk filters. Keyword fallback when vector search is unavailable.
Four report types: briefing (what to know before submitting), precedent (how the firm has done this before), risk (standards and compliance risks), prediction (what the authority having jurisdiction is likely to comment on).
CAD/BIM script generation: AutoCAD .scr scripts, AutoLISP .lsp scripts, Revit journal files, Dynamo .dyn definitions, IFC crosswalk JSON.
Simulation-Aware Systems
17 simulation-aware systems distributed across all three engines, coordinated by a central engine:
- Vanta systems (9): Lazy scheduling, spin-curvature metadata, vortex caching, chirality feedback, discrete message passing, hierarchical document contexts, memetic pattern evolution, attention-budget allocation, irreducibility detection
- Aletheia systems (6): Causal consistency for explainable findings, error correction via multi-channel parity, reality anchors cascading confidence, Merkle tree tamper detection, anomaly detection for impossible data, counterfactual logging of alternative decisions
- Daedalus systems (2): Erasure-resilient semantic storage, structural data segregation for PII and client isolation
Safety-critical content receives 4x the attention budget. Causal chains provide E&O defense by making every finding explainable. The system logs what it decided and what it considered but rejected.
See what your documents already know
Three engines. Your hardware. Your data never leaves your network. Start with an assessment.
Request an Assessment