The problem we solve
The gap is not content but access at the moment of need. A new joiner asks about a procedure that was documented years ago. A support agent opens four files to answer one question while the customer waits. Two colleagues give two different answers because each is working from a different version. Then come the reasonable objections to AI assistants: that one will invent an answer, quote a superseded policy, or surface something the asker was never cleared to see. All three are design problems. We answer them with retrieval limited to your sources, visible citations with document dates, and permission filtering applied before any passage reaches the model.
When you need this
- The same question reaches your team every week, and the answer sits in a document nobody can locate.
- New joiners spend their first weeks asking about policies that are already written down.
- Your knowledge is spread across a shared drive, email, PDFs and team chat.
- Support agents open several documents to answer a single question while the customer waits.
- You are wary of an assistant quoting a withdrawn version, or exposing a document to the wrong person.
- Your material exists in both Arabic and English, and your current search handles one of them badly.
Expected outcomes
- Answers backed by a document the reader can open and check on the spot.
- Less time lost hunting for a file or interrupting a colleague about a documented process.
- One consistent answer to recurring questions, regardless of who is asked.
- Access rights genuinely respected, so users see only what they are entitled to.
- Faster ramp-up for new staff and less dependence on a few knowledgeable individuals.
- A documented view of where your knowledge is missing or contradicts itself.
Who this is for
- Organisations with years of accumulated policies, procedures and technical documentation and no single reference.
- Support and service teams answering repeat questions where accuracy matters.
- HR functions fielding the same queries about leave, benefits and internal process.
- Sales teams that need precise product and terms information during the call itself.
- Engineering and operations businesses working from large manuals that are hard to search.
Use cases
- An internal assistant answering policy and procedure questions from approved documents.
- A support-side assistant that drafts a reply with its sources attached for the agent to check before sending.
- Natural-language search across technical manuals and operating procedures instead of keyword guessing.
- A reference for sales covering product detail and approved commercial terms.
- A customer-facing knowledge base on your site, answering strictly from published material.
- Pulling answers out of tender packs and contract documents while preparing a bid.
Deliverables
- A working assistant over your documents, with visible source citations on every answer.
- A maintainable knowledge index with a defined re-indexing path when documents change.
- A documented permission map linking each source to who may see it.
- The evaluation set and its baseline results, reusable for judging any later change.
- A knowledge gap report listing recurring questions with no supporting source.
- A short governance guide: who owns each source, when it is reviewed, how withdrawn content is removed.
Project phases
Discovery & inventory
Mapping sources, real questions and intended users, and fixing the scope of the first release.
Content preparation
Connecting sources, processing and indexing documents, handling awkward formats and Arabic text normalisation.
Assistant build
Retrieval and answering layer, enforced citation, refusal behaviour and permission filtering.
Evaluation & tuning
Running the evaluation set, reviewing answers with your subject-matter experts, and tuning retrieval.
Staged rollout
Release to a limited group inside their working channel, feedback collection, then wider rollout.
Governance & improvement
Handover of the governance guide and gap reporting, and training content owners on the refresh cycle.
The journey
Find where knowledge actually lives
We inventory the real sources and record, for each one, who owns it, how current it is, and whether it is fit to be treated as authoritative.
Start from the questions being asked
Real questions are pulled from tickets, inboxes and chat threads. They become the success criteria, instead of a general impression of what people need.
Prove it on one bounded domain
We begin with a single area that has a clear owner — HR policy, product documentation, an operations manual — so quality is demonstrated before scope grows.
Settle permissions and governance
Visibility is tied to your source systems, and we agree what is logged from user conversations and who is allowed to review it.
Put it where people already work
The assistant reaches staff inside their existing tools rather than as another window some remember and most forget.
Improve the sources, not just the model
Unanswerable questions become a ranked list of documentation gaps, so the underlying material gets better over time.
The services this solution brings together
One solution, delivered by one team across connected services — so you are not coordinating several vendors yourself.
RAG and Enterprise Knowledge Systems
Ask in plain language, get an answer from your own approved documents with the source attached.
AI Assistants and Conversational AI
A bilingual assistant that answers from your approved sources and hands off to your team with context.
APIs and Systems Integration
Documented APIs and an integration layer that keeps your systems in step — and alerts you when it cannot.
Responsible AI and Governance Advisory
A practical AI governance framework: use register, risk grading, readable policies, clear ownership.
Natural Language Processing
Classify, extract and search your written records by meaning — built for Arabic and English in practice.
Start with a package
A first step with a written scope, before you commit to the whole solution.
AI Readiness Assessment
An honest assessment of what artificial intelligence can do in your company today — and what it cannot, and why.
- Estimated timeline
- 5 to 10 weeks
- An estimate that depends on scope and how quickly reviews come back — not a contractual commitment in itself.
- Pricing
- After scoping
- Effort depends on how many sources are in scope, what condition and format they are in, and how complex your permission model and delivery channels are. We scope the first release, then review.
Scope and working details
What is in and out of scope, how we work, and what we need from you.
What is included
- Source inventory with owner and currency recorded for each, and an agreed first scope.
- Connecting the agreed sources and handling their formats, including scanned files where feasible.
- Document processing, chunking and indexing tuned for Arabic as well as English.
- Retrieval and answering layer, with citation of the supporting passage enforced.
- Permission mapping to your source systems, applied at retrieval rather than at display.
- An evaluation set of real questions with approved answers, used before launch and after every change.
- Delivery inside an agreed working channel, plus an admin view for refreshing sources.
- Training for content owners on the update cycle and the gap reports.
What is not included
- Writing the policies and procedures you are missing — we expose the gap and advise, but do not author your reference material.
- Translating an entire document archive into a second language.
- Replacing your document management system; we build on top of it, not instead of it.
- Legal or regulatory opinions derived from assistant output.
- Large-scale digitisation of paper archives, which we quote separately.
How we work
We start from questions rather than technology: what is actually being asked, who answers it today, and what that costs. Then we pick a first domain with a clear owner and material people trust, because an assistant built on contradictory documents will reproduce the contradictions faithfully. Quality is measured against a set of real questions with approved answers before anyone outside the project sees it, and measured again after every change. And the assistant is built to decline visibly when it has no supporting passage — a stated refusal is far more useful than a confident answer with nothing behind it.
What we need from you
- A named content owner per source who can confirm what is approved and what is withdrawn.
- Read access to the agreed sources and an explanation of your existing permission structure.
- A sample of genuine questions from tickets, email or team chat.
- A subject-matter expert available to review answers during evaluation.
- A decision on what user conversations are logged and how long they are kept.
Methodology and technology
The approach is retrieval-augmented generation: your documents are searched first, and the answer is composed only from the retrieved passages, with their source shown. Indexing combines semantic and keyword search so internal terminology, system names and product codes are not lost, with Arabic normalisation covering hamza forms, diacritics and spelling variation. Permissions are enforced during retrieval rather than at display, so an unauthorised passage never reaches the model. When no supporting passage exists the assistant declines and points to a human owner. Model choice and hosting are decided against your data constraints — a managed service or a deployment inside your own environment.
From our work
Lavin Fashion: a complete commerce system, built from scratch
Store, app, AI shopping assistant, automation and supplier links, for a company our owner is a partner in.
Frequently asked questions
Can the assistant make things up?
It is built to answer only from passages retrieved from your documents, to show the source alongside each answer, and to decline openly when it finds no support rather than filling the gap. No system is error-free, which is why we hold an evaluation set of real questions with approved answers and re-run it before launch and after every change — and why the source link stays visible so users can verify for themselves.
How do we stop someone seeing a document they should not?
Permissions are applied during retrieval, not at display time, so an unauthorised passage never reaches the model in the first place. We tie this to your source systems as far as their interfaces allow, rather than maintaining a second permission list that drifts out of date.
What happens when a policy is updated or withdrawn?
Sources are re-indexed on change, and a withdrawn document is removed from the index so it stops appearing in answers. Because every answer shows its source and that document date, users can immediately see whether the reference is current. The governance guide sets out who owns each source and how often it is reviewed.
Do our documents have to leave our environment?
That depends on the option you choose, and it is settled before work begins. There are architectures that run inside your own environment or dedicated cloud tenancy, and architectures that use a provider under a clear data processing agreement. We document the data path and retention either way, and nothing goes outside without your approval.
Will it answer in Arabic from English documents?
Yes. Retrieval works across both languages, so a question in Arabic can be answered from an English source. The reply comes back in the language of the question, with the citation shown in its original language so verification stays possible.
What if our documentation is incomplete or contradictory?
That surfaces quickly, and it is often the first real value clients get. You receive a gap report ranking the recurring questions with no supporting source and flagging documents that conflict. The assistant does not hide the mess — it makes it addressable.