Orcastly Starter
Free
No card needed
- 14 days with 5 users and 300 questions
- Then free forever: 1 user, 25 questions a month
- Upload PDFs, Word, Excel and scanned paper
Your company already knows the answer. Finding it is the problem.
One assistant over everything your company has written down: contracts, policies, drawings, tickets, minutes, the spreadsheet only one person understands. It answers each person from exactly what that person is already allowed to see.
Or try it yourself first: Orcastly, our platform, is free to start. Try it free
The problem
The answer exists. It is in nine places, and none of them is searchable.
Nobody is hiding anything. The knowledge is simply spread across systems that were never meant to be read together.
Where the answer actually livesA normal company, not a broken one
File sharesEmailWhatsAppThe ERPSupport ticketsPDFsScanned archiveSpreadsheetsOne person's head
The question nobody asks
A sales engineer does not know a similar quote exists, so he prices from scratch and gets it wrong. The knowledge was there; the cost of looking for it was higher than the cost of guessing.
The answer that is three years old
Someone finds the policy, the drawing, the price list. It is the old version. There is no way to tell from the file itself, and the new one is in a different folder under a different name.
The person who is the system
One engineer, one accountant, one operations manager who simply knows. The company works because of them, and stops working the week they travel.
How it works
Retrieval first. The model does not memorise your company.
The honest version of how these systems work, because the way they are usually sold makes the wrong thing sound possible.
From your documents to an answer with sourcesThe same five stages every time
1
CollectWe connect to where the documents already are: file shares, SharePoint, email, the ERP, ticketing, WhatsApp exports, scanned archives. Nothing moves to a new system your team has to start using.
2
Clean and splitDocuments are read, split into passages that make sense on their own, and tagged with where they came from, which version they are and when they were last touched. This stage decides the quality of everything after it.
3
Index with permissions attachedEvery passage carries the access rules of the document it came from. The permission travels with the text, so it cannot be lost later.
4
Retrieve, filtered before the model sees anythingA question is matched against the index, and the filter runs first. The model is only ever handed passages this person is allowed to read. It cannot leak what it was never given.
5
Answer, with the source next to itThe answer comes back with the documents it came from, so the reader can check it in one click. An answer with no source is treated as a failure, not a result.
Two different jobs, sold as oneWhat each one actually changes
Retrieval
- Changes
- What the system knows.
- Updated
- The same day a policy changes. Re-index the document and the next answer is right.
- Shows its working
- Every answer carries the passages it came from.
- Permissions
- Enforced, because the filter sits on the documents.
Fine-tuning
- Changes
- How the answer is written: your house style, your terminology, your document shapes.
- Updated
- By retraining. Weeks, and a cost, for every correction.
- Shows its working
- No. A fine-tuned claim has no source to check.
- Permissions
- None. Whatever it absorbed, it can say to anyone.
Fine-tuning has a real place here, and it is not the one it gets sold as. It teaches the model your house style, your terminology and the shape of your documents. It does not reliably teach it facts, and facts taught that way cannot be updated on the day a policy changes. Facts come from retrieval. Nobody trains a model from scratch for this, and anyone offering to is selling you the most expensive way to get a worse result.
Permissions
It answers the finance director and the intern differently, because it has to
This is the part that decides whether the project is an asset or an incident, and it is the part most proposals leave until last.
An assistant over your company's documents inherits whatever permission mess already exists. Microsoft says so about its own product: its foundational deployment guidance describes Copilot as grounding responses in "the data users already have permission to access". If a folder has been open to all staff since 2019 and nobody noticed, the assistant is what makes anyone notice — by answering a question with its contents.
So the permissions and document-hygiene audit is a paid phase of the work, not a favour at the end. We map who can reach what today, find the folders that are open to everyone by accident, and agree the rules before a single document is indexed. Companies that skip it do not get a cheaper project. They get a quieter one, until the day they do not.
Where the permission check happensBefore the model, never after
Question“What is our margin on the Alexandria contract?”Asked in Arabic or English, in chat or inside a system you already use.
FilterWho is asking, and what may they readThe access rules of every document, matched against this person. This step runs first.
ModelOnly the passages that survivedIt cannot leak what it was never handed. Nothing is filtered out of the answer afterwards.
AnswerWith the documents beside itCheckable in one click, or an honest refusal.
The same archive, four peopleIllustration of the mechanism, not a measurement
Finance director 22 / 24
Project manager 13 / 24
Sales, other region 5 / 24
New intern 1 / 24
The same question, four people"What is our margin on the Alexandria contract?"
| Finance director | Full answer, with the costing sheet and the signed contract |
| Project manager on that contract | Scope, dates and committed costs. No margin, no salaries |
| Sales, different region | That the contract exists, and who to ask. Nothing from inside it |
| New intern | Nothing. Not a summary, not a redacted version, not a hint that the question touched something |
Before you buy anything
What we will not promise you
Five things we say on the first call, because finding them out in month four is how these projects die.
No accuracy percentage.
Anyone quoting you 95% is quoting a number they cannot measure on your documents, which they have not seen. Instead we agree a golden set: 50 to 100 real questions your own experts answer first, in writing. The system is accepted or rejected against that set, in front of you. You will know exactly what it gets wrong before you roll it out.
Fine-tuning will not make it know your facts.
It changes how the answer is written, not what is true. Treating it as a memory is the single most common way these projects produce confident, expensive nonsense.
Arabic scans get tested before we price them.
A clean Arabic PDF and a photographed handwritten form are not the same problem, and neither is a 1990s scan of a typed contract. We run your actual worst documents through during discovery. If the quality is not there, you hear it then, not after signing.
Model and hosting usage is billed separately, and capped.
It is a metered cost that moves with how much the company uses the thing. Burying it inside a fixed price means someone is guessing, and it is priced so the guess cannot hurt them. You get the meter, a cap you set, and an alert before it is reached.
A wrong answer has an owner, and we write down who.
In February 2024 the British Columbia Civil Resolution Tribunal held Air Canada liable for what its own chatbot told a customer, in Moffatt v. Air Canada. The company owns what its assistant says. So the contract states which answers are advisory, which require a human to approve before they act, and what the system does when it is not sure — which is to say so.
Arabic
Arabic is the part most vendors quietly skip
It is also most of the archive in an Egyptian or Gulf company, so skipping it means skipping the point.
Mixed documents
One contract with Arabic clauses, English defined terms, and a table of figures reading the other way. Splitting that into passages without scrambling it is not automatic.
Scans and handwriting
Delivery notes, site reports, stamped approvals, twenty-year-old typed contracts. Extraction quality varies enormously, which is why we test yours before quoting rather than after.
One word, six spellings
Names, places and technical terms are written several ways across an archive, with and without diacritics and hamzas. Retrieval has to treat them as the same word or it silently misses half the documents.
Answering in the language asked
A question in Egyptian Arabic about an English document should come back in Arabic, quoting the English clause as written. Translating the clause is how a legal answer becomes wrong.
How we run it
Four phases, and you can stop after any of them
Each one ends with something you own and can judge, so the decision to continue is made on evidence rather than optimism.
How the work actually landsWeeks from start
Discovery
Pilot
Rollout
Support
0 7 14 21 28
Four phases, and you can stop after any of themHow we run it
1
Discovery1 to 3 weeksYour documents and systems mapped, the permissions audit, extraction tested on your worst Arabic scans, and the golden set written with your experts. Ends with an architecture and a price that are based on what we found, not on what we assumed.
2
Pilot4 to 8 weeksOne department, one real body of documents, working end to end with permissions enforced. Scored against the golden set in front of you. This is where you find out whether it is worth doing at all.
3
Rollout3 to 6 monthsDepartment by department, each with its own permission rules and its own golden set. Integrations, logging, an admin console and training for the people who will run it after we leave.
4
SupportOngoing, yearlyDocuments change, so the index is rebuilt and the golden set is re-run on a schedule. Model versions move; we test before we move with them. Priced as a percentage of the build, reviewed every year.
Three ways in
Start on Orcastly today. Grow into a full Enterprise Brain.
Orcastly is our own Enterprise Brain platform, built and run by WasilaDev. The engine behind every build we deliver, so you can try it on your own files before talking to anyone.
Orcastly Business
From EGP 1,000 / month
Yearly costs 10 months
- 5 to 25 users, up to 3,000 questions a month
- An assistant per department, with group permissions
- Your company's name, logo and colours
Enterprise Brain by WasilaDev
By agreement
Built with your team
- Orcastly on your own servers, or hosted by us
- Connected to your file shares, email, ERP and archives
- Permissions audit and a golden test set before launch
Common questions
Short answers
The five questions that come up on every first call.
Does our data leave the company?
That is your decision, and we build for whichever way you go. Fully on your own servers, fully in the cloud, or a hybrid where the index and documents stay inside and only the question goes out. On-premise costs more and moves slower; it is the right answer for some archives and an expensive habit for others. We give you the real trade-off in discovery rather than a house preference.
How is this different from just buying Copilot or ChatGPT Enterprise?
Those are excellent at what they cover, which is the documents inside their own ecosystem, with the permissions that ecosystem already understands. This is for the company whose knowledge is also in a file server, a scanned archive, an ERP nobody has an API for, and eleven years of email, and whose permission rules were never written down anywhere a product can read.
What happens when it does not know?
It says so, and it says what it searched. A system that guesses to look helpful is worse than no system, because it teaches people to trust it right up until the answer that matters. The golden set includes questions with no good answer, specifically to test that it refuses them.
Who owns what you build?
You own your documents, your index, your golden set, your prompts and any code we write. Where a fine-tune sits on somebody else's base model, what you can take with you depends on that model's licence, so we settle that in writing when the model is chosen rather than promising you own something that may not be portable.
How long before anyone sees a result?
The pilot is the first real answer, four to eight weeks in, and it is a working system on real documents rather than a demonstration. Anything faster than that is either a prototype on ten files or somebody else's product with your logo on it.
Where to go next
Try the live AI Lab Prompting, context, retrieval and fine-tuning, running live on a fictional company. See the systems we build CRM, ERP, portals and dashboards built around how your company already works. See automation and AI agents WhatsApp contact centre, AI agents and automations between the tools you pay for.
Tell us what you want to build.
A website, a mobile app, a CRM or an ERP: tell us what you have in mind and we'll come back with a plan, a timeline and a price. Thirty minutes is enough to tell whether we're the right team.
What do you want to build?