The YazılımAI Enterprise AI Platform
One reusable, Azure-native core that becomes your assistant, your industry solution and — where you need it — your accounting system.
This is not a generic chatbot. We start from a shared platform core that is already developed and running, then customise the interface, the content, the workflows and the knowledge architecture around your business.
One shared core, many business solutions
Every YazılımAI deployment runs the same maintained core. Your deployment is a configuration of it, not a private copy that falls behind — so improvements to the platform reach you without creating a separate product.
The shared platform core
Retrieval, assistants, live voice, conversations, lead capture, admin, security and audit — written once, maintained once.
Your configuration
Brand, tone, contact details, hand-off wording, the assistant’s rules and your knowledge base — configuration, not a code fork.
Industry editions
Hospitality, real estate and healthcare — each adding only what its sector genuinely needs.
Business modules
Reusable business domains any deployment can add. Finance & Accounting is the first.
What that means commercially: one maintained core instead of a bespoke system that ages alone, your own branding and rules without a fork, your own knowledge base, optional domain logic where you need it, and a project that starts from a working system rather than from zero.
Shared AI capabilities
Answers questions 24/7 in Turkish or English, in the language the visitor used
Uses your approved knowledge — PDF, Word, Excel, internal pages, databases and APIs
Filters out weak matches and says it does not know instead of guessing
Captures enquiries, scores them and wraps up conversations nobody closed by hand
Records what every answer was grounded in — including the answers that found nothing
Live Voice Mode: a real spoken conversation, enabled per deployment
Documents, live data, or both
Your knowledge takes one of two shapes, and the platform handles both rather than forcing everything through document search.
Documents
PDF, Word, Excel, internal pages. Retrieved passages are relevance-filtered before the model sees them, and answers carry source citations. Uploading a file is all it takes to update what the assistant knows.
Live structured data
A database, an inventory, a schedule. A plain-language request becomes an exact filter — “a one-bedroom rental in Kadıköy between 70 and 90 square metres” returns the records that actually match, cited by reference code. An exact size or budget is never approximated by a semantic search.
Grounded answers, fewer invented ones
With retrieval-augmented generation the assistant answers from your own sources rather than from model memory, and can return citations to the original documents. Passages the search engine itself rates as weak are dropped before the model sees them, so an unrelated question gets an honest “I do not have that” instead of a plausible invention. No vendor can promise zero hallucination — what we can show is where each answer came from.
Connecting a system you already run — an ERP, a CRM, a property-management or hospital system — is scoped project work rather than a switch to turn on. The data source sits behind an interface, so the assistant, the admin screens and the voice channel do not change when a real system replaces a demonstration one.
π§ Admin Agent — the assistant that checks its own work
Several AI vendors, one question, one set of your documents
An internal assistant for your own team, grounded in your internal documents rather than in a model’s general knowledge. It is part of the platform core and can be included in your deployment. What makes it different from ChatGPT, Claude or Microsoft Copilot 365 is that it is not tied to a single AI vendor: one deployment can offer models from Azure OpenAI, Microsoft Foundry, Anthropic and Google side by side, using your own credentials.
Two things a single-vendor assistant cannot do
A second opinion from a competitor
An answer written by one vendor’s model is read by another vendor’s model against the same documents, and rewritten when the reviewer finds a claim your documents do not support. The reviewer’s notes are shown, so you can see what was objected to and why. It is a second independent reading, not a guarantee.
The same question, side by side
Put one question to up to four models at once, all reading the same documents, answers in parallel columns. Because the evidence is shared, a disagreement between columns is the models disagreeing — not one of them having read something else.
Answers you can check, and costs you can see
- β Clickable citations — press a number and read the exact passage the sentence came from
- β A cited source that does not exist is refused rather than made to look real
- β Token usage per answer and a 30-day total by model — with unpriced models reported as unknown, never as zero
- β Answers stream as they are written, with a stop button that keeps what has arrived
- β Attach a PDF, Word (.docx), Excel or text file to a question — nothing is stored
- β Reads the web when asked, searching in English whatever language you ask in
π Your internal documents live in a separate index no public assistant can reach, vendor keys stay in Key Vault or your App Service configuration and never reach a browser, and a model with no key configured is never called at all.
π¨ AI Guest Concierge for Hotels
An industry edition of the platform, built per property
The same platform, tailored for hotels and tourism: a bilingual (Turkish/English) guest concierge that answers from YOUR hotel’s own knowledge base — rooms and rates, restaurants, spa, transfers and policies — with source-cited answers and a graceful hand-off to reception. Every deployment is created for one property: your brand, your documents, your rules.
- β Answers in the guest's language (Turkish / English), 24/7
- β Grounded in your hotel's documents via Azure AI Search RAG — with source citations
- β Quotes rates and availability by deterministic arithmetic over your rates and calendar — never a figure an AI model invented
- β Turns a booking wish into a consent-gated reservation REQUEST that your staff review — never a booking, a room hold or a payment
- β Fully branded per property — colors, tone, policies, knowledge base
- β Your own PMS, CRS, live rate feed and booking submission connect as scoped project work — no ready-made connector is shipped
See it in action: our fictional demonstration property “Grand Bosphorus Palace Hotel” runs this edition end to end as a public reference demo — deterministic demo rates and availability, consented reservation requests persisted in its own dedicated Azure SQL Database for a person to review, typed answers cited from Azure AI Search, and Live Voice Mode switched on there (voice is enabled per deployment, and its spoken channel is not yet at full parity with the typed concierge). It is a demonstration with synthetic data, not a customer and not a production hotel: no real booking, no room hold, no payment, and no PMS or CRS behind it.
ποΈ AI Property Advisor for Estate Agencies
Grounded in your live listings, not in documents
The same platform, for a business whose knowledge is inventory rather than paperwork. A bilingual (Turkish/English) advisor reads your live listing database: "a one-bedroom rental in Kadıköy between 70 and 90 square metres" returns the listings that actually match, each with its reference code — by text or by voice. Agents manage listings from a status board; customers search and browse on your site.
- β Understands a request in plain language — district, budget, size, rooms, rent or sale
- β Answers only from listings that exist right now, cited by reference code
- β Admin panel for listings with a status board (active, reserved, sold, rented, withdrawn)
- β The same advisor by voice — the customer can simply describe what they want
- β Your own listing system or database plugs in behind the same interface, as project work
See it in action: our fictional demonstration agency “Marmara Estates” runs this edition end to end. It is a reference implementation, not a customer — ask us for a walkthrough.
Ask our AI about the Real-Estate Editionπ₯ AI Patient Assistant for Hospitals and Clinics
The edition defined by what it refuses to say
The same platform, for the industry where the hard question is not what the assistant can answer but what it must never answer. A bilingual (Turkish/English) patient assistant handles departments, opening hours, appointment rules, check-up packages, laboratory and imaging, visiting and payment — the questions that fill a hospital switchboard — and is built to refuse to diagnose, to say how serious something is, to change a treatment, or to read a test result. On a description of chest pain or breathing difficulty it stops and directs the person to emergency care instead of asking another question.
- β Answers in the patient's language (Turkish / English), by text or by voice, 24/7
- β Grounded in YOUR hospital's own documents via Azure AI Search RAG — with source citations
- β Never diagnoses, never interprets a result, never changes a treatment — the boundary is one constant every answering mode must emit, not a paragraph that can drift
- β Emergency descriptions stop the conversation and route to emergency care, with no follow-up question
- β Collects no clinical data — no symptom field, no identity number, no file number, no result. An appointment request carries only what your staff need in order to reply to it, taken with explicit consent and under a retention window you state
- β Departments, hours and free appointment slots come from your own HIS as integration work; the request goes to your staff, and the assistant books nothing
- β In writing, deterministic code — never a second model — reads every answer before the patient does: a forbidden sentence is rewritten once and otherwise withheld in favour of a hand-off. The current Live Voice architecture is not protected by that check and relies on its realtime instruction and safety boundary instead, and we say so rather than letting the written guarantee stand for both
See it in action: our fictional demonstration campus “Altınboynuz Sağlık Kampüsü” runs this edition end to end. It is a reference implementation, not a hospital and not a customer — ask us for a walkthrough.
Ask our AI about the Healthcare Editionπ Finance & Accounting
A reusable business module, not an industry edition
A deterministic double-entry accounting and financial-management application on the same platform core, for a customer in any sector. Nothing becomes a real accounting figure until a person approves and posts it, posted entries are corrected by reversal rather than edited, and every change is recorded in an audit trail that cannot be edited or deleted.
Numbers come from deterministic financial tools. The model explains them. Code verifies every figure in the explanation before anyone reads it.
If a number in the AI’s answer cannot be traced back to a figure the accounting engine produced, the answer is regenerated once and then withheld — you get the verified figures in a table plus a notice, never an unverified sentence. No AI model verifies another AI model on this path.
- β Double-entry general ledger: draft, approve, post and reverse, with entry numbers allocated only at posting
- β Multi-company with four access levels, so the person who prepares an entry need not be the person who posts it
- β The four standard reports — Mizan, Yevmiye, Büyük Defter and Hesap Ekstresi — from one engine, so two screens cannot disagree
- β Receivables and payables with ageing, per-counterparty balances and reconciliation against the general ledger
- β Cash and bank: your book balance beside the bank’s, with the difference stated and neither side adjusted — and nothing that moves money
- β Budgets versus actuals, and deterministic cash-flow forecasting that prints its own assumptions and everything it left out
- β Import from CSV, Excel or a system you already run — always previewed, always ending at a draft a person approves
- β Document extraction proposes values; a person confirms every required field and chooses the accounts before anything reaches the ledger
- β The AI can never approve, post, reverse or pay — those capabilities do not exist in it, so no setting and no prompt can reach them
- β The accounting engine runs with no AI at all — the Copilot is an optional component you can simply not deploy
Status, stated plainly: the Finance & Accounting module is engineering-complete — built, covered by an automated test suite and validated locally against a real SQL Server engine — and it is not deployed anywhere. It is a country-neutral accounting core with no Turkish statutory compliance: no Tek Düzen certification, no e-Fatura, e-Arşiv or e-Defter, no tax or VAT filing. Statutory packages are a separate, later scope.
Ask our AI about the Finance moduleHow a request is answered
The request path below is the same in every deployment. Which orchestration engine runs it is a per-deployment choice — the platform’s own pipeline is the default, and a Microsoft Agent Framework engine can be selected by configuration alone.
Model-agnostic by design
A deployment chooses which models answer — Azure OpenAI, Microsoft Foundry, Anthropic or Google — using your own credentials, held in Key Vault. Orchestration is a per-deployment choice too: the platform’s own pipeline is the default, and a Microsoft Agent Framework engine can be selected, and rolled back, by configuration alone.
Core building blocks
- Azure OpenAI for language models, with the model list set per deployment
- Azure AI Search for hybrid keyword + vector retrieval with semantic reranking
- Azure Blob Storage for documents, with automatic chunking and embedding on upload
- Azure Document Intelligence for document extraction, where that is in scope
- A selectable orchestration engine — the platform’s own pipeline by default, Microsoft Agent Framework as an alternative
- .NET 10 and Blazor front end with an ASP.NET Core back end
- Azure SQL for business data, conversations and application state
- Azure Key Vault for secrets, with managed identity on the App Service
Key capabilities
- Knowledge segmentation by department, product line, region or customer type
- Update the knowledge base by uploading a file — no code change, no redeployment
- Conversation memory, summarisation, lead scoring and escalation rules
- Human-in-the-loop approval gates on anything with commercial or legal consequence
- Retrieval telemetry on every answer, and a live operator log of the effective configuration
Security and customer isolation
Single-tenant by design: your own App Service, database, Key Vault and search indexes, in the Azure subscription and region you choose
Your AI service key never reaches a visitor’s browser — a voice session receives only a short-lived, single-session credential
Which knowledge base a question may reach is decided server-side against a fixed list; the browser never names an index
No raw audio is ever stored — only finalized transcripts, and only where consent was recorded on the server
Encryption in transit and at rest, admin surfaces behind a separate credentialed sign-in, Key Vault for secrets, and Application Insights for monitoring
Untrusted input stays data: retrieved documents, uploads and web content are fenced and never treated as instructions, and tool or knowledge-index selection never leaves the server. Attempts to push on that boundary are counted, not obeyed.
Commercial truth is checked by code: a deterministic gate reads every public written answer before a visitor does and withholds claims the company cannot support — a certification, a price, a guarantee. It raises the floor; it is not a certificate, and we say so.
How a project runs
Discovery
goals, data sources, languages, integrations and security constraints
MVP & pilot
a working system on your own knowledge, measured with real users
Production & handover
hardening, monitoring, scripted provisioning into your Azure subscription and a documented handover
Ready to see it on your own content?
Tell us what your knowledge looks like — documents, a database, or both — and we will show you the shape of the solution.
Get in Touch