Six days became eleven minutes.

SectorEducation · Credentials

ClientUNID Solutions

Year2024

StatusProduction · Scaling

StackLLM Eval Pipeline · Document AI · pgvector + Goldset · Human-in-the-loop · Audit Trail · Next.js + Postgres

INDIMA · Education · Credentials2024

An AI-driven credential-recognition engine with a full audit trail. Six days of review become eleven minutes — verifiable, not guessed.

Challenge

A university assessed credentials by hand. In six days.

The process: PDF transcripts from 70+ countries, manual mapping onto the ECTS system, grade conversion, then an assessment with an audit trail. Eight operators. One week per case. Thousands of requests in the backlog. The tech had been attempted for years — failing on reliability, not on models.

Approach

Not a model. A loop.

We put the mapping into an eval-re-eval loop: Document AI parses the transcript, a mapping agent proposes ECTS values, a second agent checks them against a goldset and history, and a human sign-off step closes the case. Every step with an audit trail. Swapping models mid-project was no problem — because the eval pipeline measures it, not a gut feeling.

LLM Eval PipelineDocument AIpgvector + GoldsetHuman-in-the-loopAudit TrailNext.js + Postgres

Outcome

Today two operators assess what eight managed in a week.

Handling time

6 days→11 min

−99%

Operators per case

8→2

−75%

Audit findings

14 / year→0

−100%

Backlog

~2,400→<30

−98%

«We had tried three vendors. With ADLAS, the first transcript was cleanly assessed after two weeks — and after eight months it was the standard.»
Head of Recognition · University Consortium · Vienna