AURA (Beta)

Founder & Solo Engineer2026 – Present · Building in the open

AURA productizes a 5-stage AI process-audit pipeline. Auditors walk a company from a stated root problem to a measurable improvement plan — each stage is a validated LLM call with a persistent audit trail, plus a voice-first intake so interviews are transcribed straight into the workflow.

Visit the live site

The problem

Consulting-style process audits are slow, unstructured, and hard to reproduce. The upstream Python notebook proved the pipeline worked; it needed to become a real product an auditor could log in and use, with an audit trail, role-based access, and voice-first intake.

My approach

I designed and built the whole stack solo. A Fastify + Prisma + PostgreSQL backend wraps each of the 5 stages as an HTTP route, calling Azure OpenAI GPT-4o in JSON mode and validating the response against Zod schemas that mirror the notebook's Pydantic models field-for-field. A Next.js 16 frontend gates on Microsoft Entra External ID (CIAM) via MSAL; every textarea has a voice mic proxied through Whisper. Every LLM call writes an AIResponseLog row for auditability. Owner-scoped from day one — tenancy is ownership, not a separate Organization boundary.

Highlights

What I shipped on this project.

5-stage AI pipeline as HTTP routes: Decompose → Plan → Interview → Pain Graph → Solution → Measure

Azure OpenAI GPT-4o JSON mode with Zod validation and one-shot retry on schema failure

Voice intake via Azure Whisper — every textarea has a mic, transcript flows straight into the answer

Microsoft Entra External ID (CIAM) with JWKS-verified JWTs and Admin / Auditor roles

13-model Prisma schema on Azure PostgreSQL Flexible Server with atomic re-run: replaying an earlier stage wipes everything downstream

AIResponseLog audit trail — prompt version, model, duration, ok/error, raw request + response

Prompts copied byte-for-byte from the original Python notebook so behavior stays reproducible

Dev-mode bearer-token auth for local API work, hard-disabled in production

Stack

Tools and infrastructure.

Next.js 16React 19Tailwind v4Framer Motionnext-intl (EN + ES)MSAL ReactMicrosoft Entra External IDFastify 5Node 24 LTSPrisma 6Azure PostgreSQL FlexibleAzure OpenAI (GPT-4o)Azure WhisperZodPino

Outcome

5-stage AI audit pipeline productized · voice-first intake · full audit trail on every LLM call · owner-scoped by design

Previous project

Phi Tribe

Wellness Content & Community Platform

Next project

Slánú Lab (Beta)

Medical Coding & Billing Assistant

Want to build something like this?

Send a short note about what you're working on and I'll get back to you.

Start a conversation