Building a Resilient AI Audit Pipeline

LocalEyes AI converts business onboarding data, Google Business Profile signals, competitor intelligence, and ranking snapshots into structured audits and weekly improvement tasks. The hard part isn’t calling an LLM; it’s making the pipeline reliable when models fail, truncate, or return invalid JSON.

The problem

A local SEO audit depends on a chain of transformations:

1. Assemble inputs from GBP, reviews, rankings, and questionnaire data
2. Run an Analysis Pass to produce structured scores and risk flags
3. Validate output against versioned JSON contracts
4. Generate prioritized actions and seed a task pool
5. Run a Report Pass for a client-ready Markdown summary

Any broken link in that chain means a failed audit cycle and a frustrated business owner waiting for results.

Architecture decisions

We built the system as a self-contained NestJS module (~74 files, 17 REST endpoints) deliberately decoupled from legacy task systems. Each audit cycle persists:

• A SHA-256 hash of the input bundle for reproducibility
• Pipeline run metadata (stage, model, token usage, errors)
• The full universal audit record as JSON for downstream regeneration

Multi-tenant isolation is enforced at every layer: JWT + RBAC guards, with locationId scoped from trusted headers, never from request body fields.

Dual-mode AI execution

The core reliability pattern is dual-mode execution for both Analysis and Report passes:

// Simplified concept: live mode vs fallback mode
if (deepseekAvailable && !modelDisabled) {
  result = await deepseekService.analyze(input);
} else {
  result = deterministicAnalysisBuilder(input);
}

When DeepSeek is unavailable (missing API key, CI environment, or production degradation), deterministic local builders synthesize analysis from questionnaire disclosures, risk flags, and input signals. Failed cycles are marked failed with a stored error message rather than crashing the platform.

Preventing recommendation overload

Raw audit data can produce dozens of possible fixes. We implemented:

• Volume caps: 15 actions on first audit, 10 on subsequent
• Deduplication by factor ID, with negative triggers overriding duplicate score-gap actions
• A hidden task pool with weekly release capping visible work at 4 tasks per priority bucket

This converts complex audit output into digestible weekly execution plans, a product-engineering solution, not just a backend detail.

Takeaways

• Treat LLM output as untrusted input. Validate against schemas before persistence
• Design fallback paths from day one, not as an afterthought
• Version your contracts explicitly (we maintained 10+ versioned JSON specs across the pipeline)
• Separate “generate everything” from “show the client what they can act on this week”

The module is live in production at dashboard.localeyes.ai, powering audits for multi-location businesses and agencies.