Custom AI Solutions on Your Own Data
Assistants that answer from your documents and show their sources, document processing that turns forms and scans into data, and language-model features inside the software you already run. Built in your own Azure subscription or on your servers, and measured before anyone relies on them.
RAG chatbot source code
A C# Teams bot that answers from Azure AI Search
Azure OpenAI setup guide
Creating the resource and deploying models, step by step
Registered in Jodhpur since 2022
Dewiride Technologies Private Limited
What we build
Answers from your documents
Retrieval-augmented generation (RAG): the system finds the passages in your documents that answer a question, and the model writes its answer from those passages and shows where each part came from.
Document processing
Reading invoices, forms, contracts and scans, extracting the fields you need and passing them to your systems, with a person checking anything the model is unsure about. We build on Azure Document Intelligence and Azure Content Understanding, part of Microsoft’s Foundry Tools.
Language-model features in your software
Summaries, first drafts, classification and data extraction inside the applications your team already uses, rather than in a separate chat window.
Search over your own content
Keyword, vector and hybrid search on Azure AI Search, which Microsoft also markets as Foundry IQ, with semantic ranking in the Azure regions that offer it, so people find the right document even when they describe it in different words.
The data work underneath
Cleaning, splitting and indexing your content, recording who may see each document, and keeping the index current as documents are added, changed and removed. We sell source code for indexing data into Azure AI Search (opens in a new tab) on Dewiride Shop.
RAG, fine-tuning or a better prompt?
Following Microsoft’s guidance, we start with the prompt, add retrieval when answers depend on your own data, and fine-tune only when the model’s behaviour itself has to change.
| Approach | What it changes | When it fits |
|---|---|---|
| Prompt engineering | The instructions and examples sent with each request. The model itself is unchanged. | Always the first step: cheap to try and quick to change. |
| Retrieval-augmented generation (RAG) | What the model has in front of it when it answers: the relevant passages from your documents are added to each request. | Answers must reflect your private or changing information, with sources people can check. This is usually where we start. |
| Fine-tuning | The model itself, trained further on your examples. Microsoft Foundry supports supervised fine-tuning and reinforcement fine-tuning; direct preference optimisation is in preview. | The model’s behaviour, style or task performance has to change, rather than its knowledge: a fixed output format, a house style, or a smaller model trained on a larger one’s answers to do its job. |
Choosing a model
We test a few candidate models on examples of your real work and compare answer quality, cost per request and response time. Then we recommend the one that does the job well enough at a cost and speed you can accept, which need not be the largest.
In Microsoft Foundry, models sold by Azure are hosted, billed and supported by Microsoft, and they sit alongside models from partners and the community. Anthropic’s Claude models, for example, are available there as partner models, with Anthropic as the seller.
Where your prompts are processed depends on how the model is deployed. For models sold by Azure, such as Azure OpenAI, Microsoft states that Global deployments may process prompts in any Azure region, Data Zone deployments keep processing within a Microsoft-defined zone (US, EU or Asia Pacific), and Standard or Regional Provisioned deployments keep it within the Azure geography you choose. Developer test deployments carry no data-residency guarantee, and partner models such as Claude follow their own rules. If your data must stay in one place, we choose the deployment type with you and check that the model is offered there.
Models are updated and retired, so the evaluation set we build with you stays useful after launch: before you move to a new model or version, we run the set against it and compare the results with the model you use today.
How we measure quality
- An evaluation set agreed with you before launch: real questions with expected answers, including questions the system should decline.
- Automated scoring with evaluators built into Microsoft Foundry, such as groundedness (whether an answer is supported by the passages retrieved), relevance and retrieval quality. Evaluations are generally available, but Microsoft still labels some RAG evaluators, such as response completeness, as preview.
- Retrieval measured separately from the answer, so we know whether a wrong answer came from finding the wrong passage or from the model.
- A person reviewing a sample of answers, because automated scores miss things a reader notices.
- The full set run again whenever the content, the prompts or the model change.
- After launch, questions the system could not answer are logged and reviewed, and the fixes go into the content or the setup.
Your data and who can see it
We build in your own Azure subscription or on your own servers, so your documents, indexes and logs stay under your control and your contracts with Microsoft.
Microsoft states that prompts and responses sent to models sold by Azure, including Azure OpenAI, are not used to train generative AI foundation models without your permission or instruction, and are not available to OpenAI or the other providers of those models. Models from partners and the community are covered by their providers’ own terms, so we tell you which terms apply before we choose a model.
Azure AI Search can limit results to the documents each person is allowed to see. Filtering results with security filters, which match the user or group identities your application supplies, is generally available. Native enforcement based on Microsoft Entra ID and on permissions synchronised from Azure Data Lake Storage Gen2, Azure Blob Storage, SharePoint and Microsoft Purview sensitivity labels is still in preview, and a permission change reaches the search results only after the next synchronisation. So we set up access rules explicitly and test them with real accounts, rather than assume they carry over.
We follow the data-protection law that applies to your project: India’s Digital Personal Data Protection Act, and GDPR where EU personal data is involved.
Keeping running costs in check
- Microsoft bills Azure OpenAI models per token on pay-as-you-go deployments or through provisioned throughput units (PTUs), and Azure AI Search per search unit per hour on its dedicated tiers. Both appear on your own Azure bill rather than ours, and we estimate them with you before launch.
- Smaller, cheaper models for simple steps such as classification, and larger models only where they change the result.
- Tuning retrieval to pull in only the most relevant passages for each question, which keeps every model call small.
- Caching results that would otherwise be computed again for the same input.
- Spending limits agreed with you and set as budget alerts on your Azure subscription, so your team hears when costs pass them.
How we work and what it costs
Discovery call
We look at the problem, the documents and systems involved, and whether AI is the right tool at all. To begin, book a free discovery call (opens in a new tab).
Pilot on your data
A small pilot on a real sample of your documents, with the evaluation set and success criteria agreed in a statement of work.
Production build
We take what the pilot proved into production, with access rules, monitoring and cost alerts in place. When the system needs to act as well as answer, we build it as an AI agent, mainly with Microsoft Agent Framework.
Handover and support
At handover your team receives the code, including what keeps the search index up to date, plus the prompts, the evaluation set and the documentation. Afterwards we update the content, the index and the evaluation set as your documents change, as needed or on a monthly retainer, Monday to Friday, 9 AM–6 PM IST.
- Consulting
- From ₹2,500 (US$40) an hour
- Monthly retainer
- From ₹2,00,000 (US$4,000) a month
- Pilots and fixed-scope projects
- Quoted per statement of work
GST is added to Indian invoices. Rate cards by business size and add-ons are on our pricing page.
Sources
- Introduction to Azure AI Search (opens in a new tab) — Microsoft Learn
- What is Azure Document Intelligence in Foundry Tools? (opens in a new tab) — Microsoft Learn
- Microsoft Foundry fine-tuning considerations (opens in a new tab) — Microsoft Learn
- Built-in evaluators reference (opens in a new tab) — Microsoft Learn
- Data, privacy, and security for Foundry Models sold by Azure (opens in a new tab) — Microsoft Learn
- Understanding deployment types in Microsoft Foundry Models (opens in a new tab) — Microsoft Learn
Last reviewed by Jagdish Kumawat, Founder & Director
Frequently Asked Questions
Do we need to fine-tune a model?
Which model should we use?
Where does our data stay?
Does Microsoft use our documents to train its models?
How do you know the answers are right?
Can it read scanned documents and handwriting?
Can it respect who may see which documents?
What does custom AI cost?
Related services
- Microsoft Foundry & Foundry Agent ServiceProjects on Microsoft Foundry: choosing and deploying models, running agents on Foundry Agent Service, and evaluating them before and after release.
- AI Chatbot DevelopmentChat assistants that answer from your own content on your website, in Microsoft Teams or at your support desk, and pass the conversation to staff when needed.
- AI Agent DevelopmentAgents that call your systems, follow your rules and hand over to a person when they should, proved in a small pilot before they go to production.
Tell us what your agent should do
We'll tell you the simplest approach that will work — Microsoft Agent Framework, Copilot Studio or no agent at all — and what a pilot would involve.