DFMEA Agentic Review System: A Multi-Agent AI Pipeline for Automotive Safety

Why This Matters Design Failure Mode and Effects Analysis (DFMEA) is the highest-stakes engineering review process in automotive development. Missed failure modes can lead to costly post-launch recalls, regulatory penalties, or even safety crises. Today’s manual reviews are inconsistent, experience-dependent, and lack auditable reasoning.

Our solution: A multi-agent AI system on AWS that:

Identifies gaps in structural completeness, failure modes, regulatory compliance, and risk scoring. Cites evidence for every claim (ontology-grounded, schema-validated). Pauses for human approval at four critical checkpoints. Produces an immutable audit trail for regulators and engineers. System Overview The system processes DFMEA documents through a typed, ontology-driven S1→S7 pipeline:

S1 Intake: Parse and normalize DFMEA into ontology terms. S2 Decomposition: Break down parts into subcomponents. S3 Retrieval: Fetch relevant evidence (standards, historical data). S4 Analysis: Four AI specialists (Failure Mode, Structural, Regulatory, etc.) analyze in parallel. S5 Synthesis: Merge findings and resolve disagreements. S6 Risk Scoring: Assign Action Priority (AP) — safety-critical findings always rank High, unlike legacy RPN. S7 Report & Approval: Generate a visual summary + tabular report for engineer sign-off. Key Defenses Against Hallucination:

Every AI claim must cite a document, ontology entry, or standard. Outputs are JSON-LD validated against strict schemas. Four human-in-the-loop (HITL) gates ensure AI never auto-publishes. Core Architectural Principles Non-negotiable rules for all code and features:

Every claim is grounded → Citations resolve to real documents/ontology. Every artifact is typed → JSON-LD schemas enforce structure. Every action is auditable → Review ID threads through logs, metrics, and artifacts. Humans approve, AI doesn’t → Four HITL gates (ontology curation, low-confidence findings, disagreements, final sign-off). Stages are independent → Re-run any stage without reprocessing the entire pipeline. The Ontology: Our Most Critical Asset A shared graph of 11 top-level classes (e.g., ont:FailureMode, ont:Standard, ont:Material) and ~25 relations. It enables:

Multi-agent collaboration (agents speak a common vocabulary). Gap detection via part-class templates (e.g., “StructuralColumn must include outer_panel and spot_weld_array”). Reproducibility — every review records the ontology version used. Authoring Strategy:

Phase 1: Corpus-driven extraction from 10 real DFMEAs → ~200 terms. Phase 2: Import AIAG-VDA, ISO TS 16949, NHTSA vocabularies (~1,000 terms). Phase 3: Organic growth via review proposals + monthly curation. Two Decoupled Pipelines Ingestion Pipeline Land files → classify → validate → tag with provenance → dedupe → persist. Emits IngestionCompleted to trigger reviews. Review Pipeline Runs S1→S7 against ingested DFMEA. Supports re-review without re-ingestion. Why decouple?

Fix a typo in a DFMEA without reprocessing all reviews. Replay reviews with updated ontology/data. Human-in-the-Loop (HITL) Gates Gate Trigger Actor SLA Ontology curation New term proposed Data Scientist 2 business days Low-confidence findings S4 confidence < 0.7 Reviewer 4 hours Agent disagreements Contradictory S5 assertions Reviewer 1 business day Final report approval S7 report ready Engineer 5 business days Decisions feed an eval dataset for continuous model improvement.

AWS Service Inventory (Phase 1) Core: Bedrock AgentCore, Step Functions, EventBridge, SQS/SNS. Data: S3, DynamoDB, Bedrock KB, Aurora (deferred). Ops: CloudWatch, X-Ray, CloudTrail, KMS, A2I for HITL. UI: Amplify, Cognito, API Gateway. Phase 2 adds: Neptune (ontology graph), OpenSearch, Textract, GuardDuty.

Phased Rollout (6–7 Weeks) Foundation Prep: AWS accounts, budget alarms, repo setup. Content Authoring: Seed ontology, synthetic DFMEAs, golden eval set. Skeleton Deployment: CDK stacks, data layer stubs, UI shell. Agent Implementation: Intake Lambda, four specialists, HITL gates 1 & 4. Polish & Demo: Engineer UI, audit package endpoint, cost-per-review metrics. Risks & Mitigations Risk Mitigation Bedrock model access delays Request day 0 in two regions. Ontology authoring slows progress Allocate domain-expert hours; Marcus owns Phase 1. Synthetic data lacks realism Domain experts review templated DFMEAs. Cost overruns in agent testing Hard budget: $200/mo; use Haiku-first. How to Contribute Senior Engineers: Own orchestration, agents, ingestion, or UI. Data Scientists: Author ontology, build eval sets, tune prompts. Domain Experts: Validate AIAG-VDA conformance, review findings. SRE/DevOps: Harden CDK, build dashboards, cost instrumentation. Communication Norms:

Architecture decisions → docs/decisions.md. Operational changes → PR review. Disagree-and-commit: Surface concerns in writing, then move on.