profilo-studio

Made by ariya-tech.net

Case Studies

Systems in production, measured.

Four we can talk about, with the before/after numbers and a research note on what actually moved them. Names withheld where the contract asks.

Government services01 / 04

Bilingual citizen assistant for a national portal

Challenge

A services portal was routing every question — however routine — to a human queue, with a two-day median reply and no Arabic parity.

Approach

A retrieval assistant grounded in the regulations, with a re-ranker fine-tuned on real tickets and a refusal path that escalates cleanly when confidence drops.

ResultsBeforeAfter
Ticket deflection0%62%
Median reply2 days340ms
Unsupported answers0.9%

Research note: most of the accuracy came from the re-ranker, not a bigger model. Swapping the 70B for a fine-tuned 8B cut cost 83% with no measurable quality loss.

Banking02 / 04

Fraud-triage agents for a retail bank

Challenge

Analysts were reading every flagged transaction by hand; the backlog meant genuine fraud sat unreviewed for hours.

Approach

A triage agent that gathers context from four systems, drafts a recommendation with its reasoning, and hands the analyst a decision instead of a raw alert. The human still signs.

ResultsBeforeAfter
Time to review4.2 hrs9 min
Analyst throughput5.8×
False-positive rate31%11%

Research note: the win was the trace. Because every recommendation cites its evidence, analysts trusted it in week one instead of month three — adoption, not accuracy, was the real bottleneck.

Energy03 / 04

Document intelligence over engineering archives

Challenge

Forty years of standards, drawings, and inspection reports — searchable only by filename. Engineers rebuilt answers that already existed.

Approach

A retrieval platform over the full archive with table- and diagram-aware parsing, so an answer can cite a clause and the figure beside it. Deployed inside the client’s own region.

ResultsBeforeAfter
Time to find a standard35 min40s
Answers with citation100%
Archive coverage18%96%

Research note: diagram-aware chunking mattered more than embedding choice. Parsing tables as structure, not prose, doubled retrieval precision on inspection reports.

Healthcare04 / 04

Arabic clinical-coding assistant

Challenge

Coders were mapping free-text Arabic notes to billing codes by memory, with a review backlog and inconsistent results between shifts.

Approach

A fine-tuned model that proposes codes with a confidence and the sentence it relied on, leaving the coder to confirm rather than recall. Evaluated against a physician-adjudicated set.

ResultsBeforeAfter
Coding accuracy82%96.5%
Notes per hour1439
Inter-shift variancehighlow

Research note: showing the source sentence changed behaviour. Confidence scores alone were ignored; a highlighted span was trusted — the interface was the alignment.

Tell us what you want to ship.

A short brief is enough to start. We reply within two working days with a scoping call — no forms behind the form.

Start a project

Mirage AI

Applied-AI studio in Masdar City, Abu Dhabi. We take models to production and stay on-call for them.

The studio

Hours

Sun – Thu · 09:00 – 18:00 GST · On-call for production 24/7

© 2026 Mirage AI — All rights reserved.

From weights to production

012Mirage AI