S/11 · Services
AI solutions
AI that does real work, and shows its working.
Assistants, agents, smart search, vision, speech and forecasting — built into your product and processes, measured against real results and kept under control.
- Models
- Hosted or self-hosted
- Measured with
- Evaluations on your own data
- Built with
- Guardrails & human review
- 01
People spend their days reading documents and retyping them.
- 02
Customers can’t find answers that already exist somewhere.
- 03
Decisions are made on gut feeling while the data sits unused.
- 04
You tried an AI demo and it fell apart on real data.
- D/01
Assistants & agents
An assistant that answers from your own policies and records, says where each answer came from, and can act through your systems: look up an order, open a ticket, draft a reply. You decide what it may do alone and what needs a person’s confirmation first.
Included
- Answers grounded in your documents, with the source cited
- Actions through your APIs, each one allowed by name
- Confirmation before anything that changes data or costs money
- A hand-over to a person when it isn’t sure
support assistant · example-shop.com Can a customer return a kettle after 20 days?
Yes. Unused items can be returned within 30 days, in their box. Source: Returns policy, section 2.1.
Then start a return for order A-2291.
Order A-2291: one kettle, delivered 14 days ago. Create the return and email the label to the customer? This needs your confirmation.
- D/02
Smart search
Your documents are split into passages and indexed by meaning, so a question finds the right paragraph even when it uses different words. The best passages go to the model, which answers from them and nothing else, and links back to where it read it.
Included
- Indexing of your files, wiki, tickets or database
- Access rules kept: people only find what they may see
- Answers that quote and link their sources
- Re-indexing on a schedule, or whenever a document changes
docs/search/rag-pipeline.mmd - D/03
Document AI
Scans and PDFs are read into fields your systems can use: supplier, dates, totals, line items. Every field carries a confidence score and is checked against rules, so clean documents go straight through and doubtful ones wait for a person.
Included
- Extraction of the fields you need, in Persian and English
- Checks such as totals that add up and known suppliers
- A review queue for anything below the confidence threshold
- Results sent to your accounting system, database or sheet
inbox/invoice-1187.pdf · extracted Field Value Confidence Check Supplier Example Paper Co. 0.99 known supplier Invoice no. INV-20417 0.98 not a duplicate Date 2026-10-28 0.97 ok Subtotal 1,240.00 0.96 matches lines Tax 111.60 0.71 sent to review - D/04
Vision & speech
Calls, voice notes, photos and camera feeds become data you can search and count. A support call becomes a transcript with its reason and order number; a product photo becomes a pass or a flagged defect. Anything the model is unsure of goes to a person.
Included
- Speech-to-text tuned and tested on Persian audio
- Tags pulled from the content: intent, order numbers, defects
- Low-quality input detected and routed for review
- Batch runs or live processing, as the job requires
pipelines/call-transcripts.log $ transcribe calls/2026-10-28/*.wav --lang fa
call-0412.wav 3m12s 2 speakers
intent: refund request order: A-2291
call-0413.wav 1m48s 2 speakers
intent: late delivery sentiment: negative
! call-0415.wav poor audio → human review
✓ 38 calls · 36 tagged · 2 sent to review
- D/05
Forecasting
We build forecasts from your own history: sales, bookings, cancellations. Each one comes with a range, so you plan for the likely case and know how wide the unlikely ones are. We test every model on past periods it has never seen before you rely on it.
Included
- A forecast for each product, branch or segment you need
- A range around every number, not a single guess
- Back-testing on past periods, with the error shown
- Forecasts refreshed on a schedule and delivered where you plan
forecasts/demand · SKU 1042 Week Forecast 80% range Last year Week 45 420 units 380–465 391 Week 46 455 units 405–510 402 Week 47 610 units 520–700 577 Week 48 530 units 450–620 498 - D/06
Evaluation & guardrails
Before launch we build a test set from your real cases and score the system against it, and every change of prompt or model is scored again. Guardrails keep it on topic, keep private data out of answers, and stop it acting outside its limits.
Included
- A test set of real questions and documents, with expected answers
- An evaluation run on every change, with results compared
- Checks for personal data, off-topic requests and prompt injection
- Cost and response time tracked next to accuracy
evals/support-assistant · run 14 Correct answers
91%
+4 since run 12
Sources cited correctly
96%
212 of 220
Out-of-scope refused
24/24
no change
Cost per answer
$0.004
median 1.9 s
- 01
Find the job
We pick a task where AI saves real time or money, and define what good looks like.
- 02
Prove it
A prototype measured on your real data before anything is promised.
- 03
Build it in
Integrated with your systems, with fallbacks and human review where it matters.
- 04
Monitor
Accuracy, cost and drift tracked in production, so it stays good.
A typical AI project takes six to ten weeks from the first workshop to a system in daily use. You see measured results on your own data by week three, before most of the building starts.
- Week 1
We choose the task with you, collect real examples, and agree what a good result looks like.
- Weeks 2–3
A prototype and a test set, with the first scores on your own data.
- Weeks 4–6
The system built into your tools, with guardrails, fallbacks and a review queue.
- Weeks 7–8
A pilot with real users, and fixes based on what the evaluations and the users find.
- Weeks 9–10
Launch, monitoring dashboards, and a handover to your team.
The code
The full source in your repository: integrations, pipelines and the services around the model.
Prompts and configuration
Every prompt, model setting and guardrail, versioned and explained.
The evaluation set
Your test cases and the scripts that score them, to check any future change or model.
Your data and indexes
Indexes, extracted data and any trained models stay on your infrastructure or accounts.
Monitoring dashboards
Accuracy, cost, response time and drift, visible to your team.
A review after launch
A check a month in on how it performs in real use, and what to tune.
- LLMs, hosted & self-hosted
- RAG
- Vector search
- OCR & vision
- Speech-to-text
- Evaluations
- Guardrails
- Forecasting
Does our data leave our servers?
It doesn’t have to. We can run models on your own infrastructure when privacy requires it.
What if the AI gets it wrong?
We measure how often it does, design for it with confidence thresholds and human review, and keep improving it.
Does it work in Persian?
Yes. We test on Persian data specifically, because quality varies between models and languages.
What do you need from us to start?
Real examples of the work: a few hundred documents, questions or records, and someone who can tell us what a correct answer looks like.
What does it cost to run?
It depends on the volume and the model. We measure the cost per task in the prototype, so you know the running cost before you commit to building.
Tell us what you are building.
Every project starts with a consultation: you bring the idea and everything you know, we turn it into a design you can build on.
Book a consultation