CASE STUDY

An AI audit platform with a human on every send

Audit season buries finance teams in document requests and queries. Oditable ingests a company's financial documents, drafts responses to auditor queries with citations to the evidence, and refuses to send anything until a human approves it. We built the whole product: three apps, the AI pipeline and the marketing site.

Client
Oditable
Sector
AI development
Relationship
2026
Services
Product design · SaaS development · AI pipeline engineering · Marketing site design and build

3

Apps in one platform

1,857

Automated tests

6

User roles

100%

Human-approved

Oditable marketing site showcase

01 / the problem

Audit responses are slow, manual and high-stakes.

Every audit query means finding documents, assembling evidence and writing a defensible response. Get it wrong and the audit drags. The founder, a chartered accountant, wanted AI to do the heavy lifting without ever letting it speak to an auditor unsupervised.

That constraint shaped the architecture: the AI can only draft from cited evidence, confidence is scored on every draft, and the approval gate is enforced in the database itself, not just the interface.

  • 01 Cut the manual slog out of audit responses
  • 02 Keep a human approval on everything that leaves the building
  • 03 Make every AI claim traceable to a source document
  • 04 Isolate each customer's financial data completely
Oditable audit query board showcase
Oditable ai-drafted response showcase

02 / the solution

Drafted by AI. Approved by people. Enforced by the database.

Next.js 16PostgreSQL + pgvectorClaude + Voyage + CohereDrizzleBullMQAstro marketing site

Documents flow through a cascading extraction pipeline, get classified and chunked, then indexed with finance-tuned embeddings. When an auditor raises a query, hybrid search plus a reranker pulls the best evidence and Claude drafts a response constrained to what it can cite, with a confidence score deciding whether it's even worth showing.

Around that pipeline we built the product: a client dashboard with a Kanban of every query from draft to sent, a separate portal where auditors submit and track requests, a super-admin app for running the SaaS, and a warm, editorial marketing site that doesn't look like accounting software.

what we delivered

The work behind the result

Document ingestion and AI classification

The detail work that lifts the product above its spec.

Hybrid retrieval with reranking

The detail work that lifts the product above its spec.

AI-drafted responses with citations

The detail work that lifts the product above its spec.

Confidence scoring and approval gates

The detail work that lifts the product above its spec.

Separate client, auditor and admin apps

Mapping the users, journeys and decisions before anything was designed.

Schema-per-tenant data isolation

The numbers that matter, visible while they can still change decisions.

Stripe subscription billing

Money moving safely, from checkout through to settlement.

Astro marketing site with persona pages

The detail work that lifts the product above its spec.

03 / the result

AI speed with audit-grade control.

Oditable went from discovery to a functionally complete, QA'd platform in five months of build: 184,000 lines of TypeScript, 1,857 automated tests and an AI pipeline that never sends unsupervised.

Next case study

OI Group

start a project

Need your digital product to do more than look good?

We'll map the journey, the systems behind it and the result the product needs to produce.