case study · Microservices · RAG

CertiFlow

Safety compliance review, with an AI that cites OSHA.

Role
Solo: architecture, backend, frontend
Timeline
2026
Status
Not live
Links
live ↗code ↗
Stack
SvelteKitNode.jsTypeScriptBullMQRedisPrismaPostgreSQLGeminiRAGDocker

Site teams upload field reports, photos, logs and incident documents; an audit pipeline compares them against OSHA construction standards and an AI agent answers safety questions grounded in the regulations.

01

The problem

Construction safety compliance runs on spreadsheets, email threads and binders. Inspections get done, but the evidence lives in ten places, and ‘are we compliant with this standard?’ depends on who you ask.

Who it’s for

Compliance and safety teams on construction sites.

02

Goals & constraints

Goals

  • One pipeline from upload to audit.
  • Automate evidence classification.
  • An assistant that answers from the actual regulations.

Constraints

  • Wrong answers about safety are worse than no answer, so the AI had to be grounded and citeable.
  • Business rules change often, so they had to stay isolated from plumbing.

03

Approach

  1. 1.

    API gateway + independent services

    An API gateway in front of a compliance service and an AI worker, so audit logic and AI processing scale and deploy separately.

  2. 2.

    Domain-driven core

    Compliance rules sit in their own domain layer, separate from transport and storage.

  3. 3.

    Event-driven evidence pipeline

    Uploads emit jobs on BullMQ/Redis; the AI worker classifies and audits them asynchronously.

04

Architecture

Microservices, DDD, event-driven queues and a RAG agent, containerised with Docker.

  1. 01SvelteKit UI → API gateway
  2. 02compliance service ⇄ domain layer ⇄ Prisma / Postgres
  3. 03upload → BullMQ (Redis) → AI worker
  4. 04question → RAG agent → OSHA standard chunks → grounded answer

05

Key decisions

RAG over fine-tuning

Standards get updated. Retrieval means updating a document store, not retraining, and every answer can point back to its source.

Queue-first evidence handling

Classification is slow and failure-prone. Doing it off the request path keeps the UI fast and makes retries free.

!

Challenges

Docker on Windows

Prisma binary-target mismatches, Alpine OpenSSL conflicts and BuildKit DNS failures, each solved and written down.

Chunking regulations well

Retrieval quality depended far more on how standards were chunked than on the model.

✓

Outcome & lessons

  • Working end-to-end system shown on video; hosted instance currently offline.
  • My first build combining microservices, DDD, event-driven design and RAG.

What I learned

  • Grounding is a product feature, not an implementation detail.
  • DDD pays for itself the first time a rule changes.

What’s next

  • Bring the hosted demo back online
  • Inline citations in agent answers
next project24Flow→