FT
Fine-tuning & alignment
Supervised fine-tuning, preference optimisation, and Arabic-first evaluation sets — so the model speaks your domain, not the internet average.
profilo-studio
Made by ariya-tech.net
Applied AI Studio · Abu Dhabi
From weightsto production.
A model on a benchmark is not a product. We take open and frontier models the rest of the way — fine-tuned, grounded in your data, wrapped in agents, and held to the metrics that matter to you.
99.95%Measured uptime
340msP95 inference
12Systems in production
2021Studio since
// Capabilities
What we build
Four disciplines, one delivery team. We do not hand a model over the wall — we own it through to the on-call rotation.
AG
Agent systems
Tool-using agents with typed interfaces, guardrails, and traces you can read. Bounded autonomy, never a black box.
RAG
Retrieval platforms
Hybrid retrieval, re-ranking, and citation-grounded answers over your documents — bilingual, permission-aware, and evaluated per query.
OPS
Deployment & ops
Inference that meets a latency budget, observability on every call, and rollbacks that take seconds. We carry the pager.
// Benchmarks
Measured, not claimed
A representative production deployment — an Arabic support assistant grounded in a client knowledge base. Baseline is the off-the-shelf model; Mirage is the shipped system.
Held-out evaluation set, 1,400 queries, Arabic and English. Q1 2026.
| Metric | Baseline | Mirage | Δ |
|---|---|---|---|
| Answer accuracy | 71.2% | 94.6% | +23.4 |
| P95 latency | 1,280ms | 340ms | −73% |
| Cost / 1M tokens | $18.40 | $3.10 | −83% |
| Unsupported claims | 8.1% | 0.9% | −7.2pts |
// Pipeline
How a request flows
Every agent we ship runs the same disciplined loop. Retrieval grounds it, reasoning plans it, tools act, and an evaluator scores the result before it reaches a user.
- 1
Ingest
Parse & classify the request
- 2
Retrieve
Hybrid search over your data
- 3
Reason
Plan steps, pick tools
- 4
Act
Call tools, draft the answer
- 5
Evaluate
Score, cite, or escalate
Featured case · Government services
An Arabic assistant that answers 62% of tickets without a human.
A bilingual retrieval assistant for a national services portal — deflecting routine questions, citing the source regulation on every answer, and escalating cleanly when it is unsure.
See the full study62%
Ticket deflection
340ms
Median response
0.9%
Unsupported answers
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