Technologies
AI models
We work with the major commercial models, Anthropic Claude, OpenAI GPT and Google Gemini, and with open weight models such as Llama, Qwen, Mistral and DeepSeek. They run on Azure AI Foundry, AWS Bedrock, Google Vertex AI or your own servers, depending on where your data is allowed to go. We pick the model per task after testing it on your own documents.
3 tools
Commercial models
Claude
Long documents, careful reasoning, coding and agent work.
OpenAI GPT
General assistants, structured extraction and voice features.
Google Gemini
Multimodal input such as scanned forms, images and video.
4 tools
Open weight models
Llama
Self hosted assistants where data must stay on your servers.
Mistral
Efficient models for classification, routing and summaries.
Qwen
Strong multilingual and coding performance at modest sizes.
DeepSeek
Reasoning tasks on private infrastructure at lower cost.
4 tools
Hosting and tuning
Hugging Face
Finding, evaluating and fine tuning open models and embeddings.
PyTorch
Fine tuning and custom models for forecasting or classification.
vLLM
Serving open models on GPUs with good throughput.
Ollama
Running models locally for development and small on premise setups.
What we build with it
Typical work in this stack
- Document extraction
- Reading invoices, purchase orders, KYC papers and lab reports into structured fields, with low confidence items sent for review.
- Private assistants
- Chat over your policies, manuals and past tickets, with answers that cite the source page.
- Classification and routing
- Sorting emails, tickets and WhatsApp messages by intent and urgency so the right team picks them up.
- Summaries and drafting
- Call notes, case summaries and first drafts of replies or reports for a person to check and send.
- Self hosted deployments
- Open models on your own servers or private cloud when data cannot leave your network.
How we choose
We build a small test set from your real documents and questions, then run two or three candidate models against it and compare accuracy, speed and cost per thousand requests. Often a smaller, cheaper model is good enough for routine steps, with a larger model reserved for the hard cases. Data rules come first: if personal or regulated data cannot leave India or your network, that narrows the options before quality is even considered. We keep the model behind a thin layer in the code, so switching provider later is a configuration change.
Questions
Good to know
Other parts of the stack
Is our data used to train the provider's models?
Not on the enterprise platforms we use, such as Azure AI Foundry, AWS Bedrock and Vertex AI. We check the terms for each project and can self host if needed.
Which model is the best?
It depends on the task, language and budget. We test on your own data and show you the results before recommending one.
Can models work in Hindi or Gujarati?
The larger models handle Hindi and Gujarati reasonably well. We measure accuracy in your languages before agreeing targets.
Start a conversation
Tell us what you are trying to build.
Share the problem in a short call. You leave with a clear view of the approach, the team and the cost.
A specialist replies within one working day. The first conversation is free.