AI Should Explain Engineering Validation, Not Invent It
May 2026 · Damir Andrijanic · 9 min read
The high-stakes validation trap
Most AI applications in 2026 focus on generation: writing code, drafting emails, or producing generic text. While impressive, this generative model falls apart in high-precision engineering workflows like BIM (Building Information Modeling). In CAD/BIM validation, if a compliance reviewer asks if a structural room meets safety codes, a 'hallucinated' or plausible-sounding answer is worse than no answer.
BIM models represent complex spatial elements—levels, rooms, doors, and fire boundaries. Traditional compliance checks are done via strict, deterministic formulas. The moment we inject standard generative AI into this loop, we risk hallucinating metrics, mapping wrong relations, or declaring compliance where it fails. Yet, simple deterministic error codes are cryptic and hard for human architects to review. How do we get the best of both worlds?
Constrained Explanation: The ADAS Spec Validator
I built the ADAS Spec Validator to prove a fundamental design pattern: AI should explain engineering validation, never invent it.
The system splits validation into two isolated layers. First, a client-side deterministic rule engine, built with TypeScript and Zod, parses Revit/CAD JSON models and evaluates strict compliance requirements (e.g. minimum room areas, fire door connectivity, path distances). It emits a strict evidence contract. Second, an AI explanation router (Gemini 2.5 Flash / OpenAI) receives ONLY this verified evidence. It is strictly forbidden by an ADAS System Prompt from inferring missing facts or inventing data, focusing solely on explaining the exact reasoning in natural language.
System Architecture & Data Flow
The data flow guarantees complete separation of the AI explainer from raw model geometry. Mathematical calculations and boundaries are executed deterministically.
Core engineering stack
BIM Extractor Boundary
A C# .NET console application modeling the Revit API boundary, extracting levels, rooms, and door relations into a sanitized JSON contract.
Type-Safe Validation
TypeScript and Zod schemas to ingest, validate, and sanitize the normalized CAD/BIM data model, preventing garbage-in-garbage-out errors.
Deterministic Rules
Strict TypeScript rule engines evaluating custom compliance rules (like MinimumRoomAreaRule, MinimumDoorWidthForRoomTypeRule) and producing detailed validation evidence.
AI Explanation Engine
Google Gemini 2.5 Flash API as a fast, high-context explanation engine, with automated provider fallbacks to OpenAI's GPT-4o-mini.
Reliability Testing
Vitest suite verifying type safety, rule engine correctness, fallback mechanics, and strict prompt compliance boundaries.
Visual UI Interface
Next.js 15 (App Router) and Tailwind CSS displaying real-time validation passes, violations, and the natural language explanation panel.
Key architectural breakthroughs
1) Eliminating LLM Hallucinations in Compliance
The Risk: When an LLM is directly asked to read a raw BIM JSON model and assess safety rules, it frequently confuses IDs, fabricates wall thicknesses, and miscalculates simple geometry.
The Solution: I isolated the AI entirely from the raw calculation. The TypeScript rule engine executes the mathematics, compiling a structured array of 'evidence' (observed value, expected threshold, target IDs). The AI is presented with this evidence and a system prompt that mandates returning 'I cannot determine that from the available model evidence' if any fact is missing.
2) Handling Missing and Unknown States Gracefully
The Risk: Real-world CAD exports are incomplete. Missing areas, unmapped doors, or missing level tags usually cause standard rules to crash or AI models to invent missing data.
The Solution: I implemented a three-state validation model: PASS, FAIL, and UNKNOWN. If a room has no area parameter, the rule engine marks it UNKNOWN and outputs precise evidence describing the missing property. The AI respects this state, explaining exactly why the verification is stalled rather than trying to guess it.
3) C# Revit Extraction Normalization
The Risk: Revit objects contain thousands of raw parameters, cyclic structural dependencies, and internal coordinates that are too noisy and heavy for web apps or prompt contexts.
The Solution: I built a C# console extractor that resolves these heavy dependencies using a Revit-inspired DTO structure. It flattens spatial elements into standard Levels, Rooms, and Doors JSON, representing the exact data contract between desktop CAD and web compliance engines.
The UX: Bringing Clarity to Complexity
Human reviewers don't want black boxes. The ADAS interface pairs clean status badges (PASS / FAIL / WARNING) with direct highlights of affected element IDs (e.g. rm-meet-01). The AI explanation acts as a smart, natural-language sidekick—brief, evidence-backed, and directly telling the architect which door parameter to fix in Revit.
Proven Results
By separating deterministic evaluation from prompt-based translation, the system achieves 100% mathematical precision while maintaining natural user interaction.
Validation Accuracy
100% Deterministic
AI Explanation Logic
Evidence-Grounded
This prototype shows that engineering-grade trust is achievable when we treat the AI as a communication interface rather than a calculation engine.
Takeaway
Don't let AI calculate what a 10-line math formula can solve perfectly. Use code for math, logic, and safety guardrails. Use AI for context, translation, and user-centric explanation. That is the architecture of trust.
BIM Compliance & Validation Studio
Experience the evidence-constrained AI system and real-time floor plan editor directly inside your browser. Drag and drop model files, edit sizes live on the floor plan, define specs, and query the cognitive terminal.
ADAS Spec Validator
Deterministic compliance engine mapping Revit / AutoCAD models against strict architectural requirements.
Storey Compliance
6 / 11 Passed
Deterministic requirements satisfied
6
2
3
CAD / BIM Spatial Map
Interactive visual rendering of model boundaries and corridors
Interactive Model Floor Plan
Click elements to inspect; validation errors glow dynamically.
BIM / CAD Extract Connector
Drag JSON or .ifc files to ingest Revit boundaries locally.
Upload Model (JSON/IFC)
Extract elements from IFC boundary or Zod JSON files.
Upload Rule JSON
Inject raw JSON file outlining validation rule arrays.
No-Code Spec Builder
Visually build and inject compliance rules directly into the deterministic validator.
Preset templates
Engine Validation Log11 specs checked
Deterministic rules evaluate parameters instantly with strict evidence outputs.
Stockroom A satisfies minimum area (18.2 sqm >= 15 sqm).
Rule: Stockrooms must be at least 15 sqm
Technical Evidence Fact
Room area was compared against deterministic threshold.
Cannot verify area for Stockroom B; area is missing.
Rule: Stockrooms must be at least 15 sqm
Technical Evidence Fact
Area parameter is not available in normalized model data.
Office 101 satisfies minimum area (9.4 sqm >= 8 sqm).
Rule: Offices must be at least 8 sqm
Technical Evidence Fact
Room area was compared against deterministic threshold.
Office 102 violates minimum area (6.1 sqm < 8 sqm).
Rule: Offices must be at least 8 sqm
Technical Evidence Fact
Room area was compared against deterministic threshold.
Stockroom A has door widths below 0.85m.
Rule: Stockroom doors must be at least 0.85m wide
Technical Evidence Fact
Connected door width is below deterministic threshold.
Cannot verify door widths for Stockroom B; door relationship data is missing.
Rule: Stockroom doors must be at least 0.85m wide
Technical Evidence Fact
Connected door identifiers are required but not available.
Stockroom A has at least one connected door.
Rule: Every room must have at least one connected door
Technical Evidence Fact
Room-to-door relationships were checked for minimum connectivity.
Office 101 has at least one connected door.
Rule: Every room must have at least one connected door
Technical Evidence Fact
Room-to-door relationships were checked for minimum connectivity.
Office 102 has at least one connected door.
Rule: Every room must have at least one connected door
Technical Evidence Fact
Room-to-door relationships were checked for minimum connectivity.
Meeting Room A has at least one connected door.
Rule: Every room must have at least one connected door
Technical Evidence Fact
Room-to-door relationships were checked for minimum connectivity.
Cannot verify connected doors for Stockroom B; relationship data is missing.
Rule: Every room must have at least one connected door
Technical Evidence Fact
Connected door relationship is missing in normalized data.
BIM Model Studio
No element selected
Select any room or door on the floor plan to inspect details and edit parameters in real-time.
Model Builder
ADAS AI COGNITIVE CORE
Evidence-Constrained Agent
The chatbot is strictly bounded to the generated deterministic rule outcomes. No hallucinated rules.
Agent Persona
Preset Queries
Deterministic Compliance Policy
The chatbot answers are strictly grounded in active spatial facts. Missing properties default to `unknown`, eliminating any AI-hallucinated certification passes.
Explore the underlying prototype source code and design patterns in my portfolio repository or download the schema contract: