What we build
AI integration
AI features added to the software you already run. Document extraction, search over your own data, assistants and summaries, built into your product or internal systems with Claude, Gemini or OpenAI models. Accepted against test cases you agree in advance.
The situation
A demo is easy, a dependable feature is not
Getting a language model to answer a question in a demo takes an afternoon. Getting it to answer your customers' questions correctly, from your own documents, at a cost per request you can live with, is real engineering work. Teams usually find this out after the demo: answers that sound right but are wrong, invoices read with the GST number in the wrong field, costs that climb with usage. An AI feature needs the same discipline as any other feature: a clear job, test cases, and a plan for what happens when the model is unsure.
What we build
AI integration, in practice.
Search and answers over your own data
Retrieval augmented generation over manuals, policies, tickets or contracts, with answers that cite the source passage so people can check them.
Document reading and extraction
Invoices, purchase orders, delivery notes, KYC papers and forms read into structured fields, with low confidence items sent to a person.
Assistants inside your product
Chat or command features that act on your users' own records, limited to what each user is allowed to see.
Summaries and drafting
Summaries of calls, tickets and long threads, and first drafts of replies or reports that a person edits and sends.
Classification and routing
Incoming emails, support tickets or WhatsApp messages tagged and sent to the right queue, with a confidence score you can audit.
Model choice, evaluation and cost control
Testing Claude, Gemini, OpenAI and open models against your own examples, then picking on accuracy, speed and cost per request.
Typical sizes
Roughly how many hours.
Planning sizes only. The real number comes from a signed scope after the first conversation.
150HOURS
One focused AI feature
A single job such as invoice extraction or a summary feature, with an evaluation set and a fallback. Roughly INR 29,850 before tax at INR 199 an hour.
180TO400HOURS
Search and answers over your documents
Ingestion, a vector store, an answer interface with citations, access rules and an evaluation set. Roughly INR 35,820 to INR 79,600 before tax.
450TO900HOURS
Several AI features across a product
A set of features sharing one model layer, monitoring and cost tracking. Roughly INR 89,550 to INR 1,79,100 before tax.
An example
Illustrative: a Bengaluru SaaS team adds document extraction
Suppose a Bengaluru SaaS company sells accounts payable software and its customers want supplier invoices read automatically. The specification sets out the fields to extract, including GSTIN, invoice number, line items and tax amounts, and attaches a set of real sample invoices with the correct answers. Acceptance is tested against that set, with an agreed rule that uncertain fields are flagged for a person to check. Model usage costs are agreed in writing before work starts.
Technology we often use
- Anthropic Claude
- Google Gemini
- OpenAI
- Python
- FastAPI
- LangChain
- PostgreSQL with pgvector
- TypeScript
- AWS
Chosen to fit the project and your team, and written into the specification.
Questions
About ai integration.
How can you accept AI work when the answers vary?
By agreeing a test set before the build: real examples with the right answers, and a rule for what counts as a pass. The specification can also say how the feature behaves when it is unsure. You test against that set before you accept.
Which model will you use?
The one that does your job best at a cost you accept. Anthropic Claude, Google Gemini and OpenAI models are all common choices, and open models suit some cases. If you would prefer a hosted builder to custom code, AI apps can also be built on AksharAI (aksharai.com), a Code Colonies platform.
Is our data sent to the model provider?
Usually some of it is, so this is decided early. The specification records which provider and region is used, what data is sent, and what is masked or kept out. If personal data of Indian users is involved, the DPDP Act applies; confirm your obligations with your lawyer.
Related
Related work and reading
Start with a conversation
Tell us what you want built. We will tell you plainly whether it fits.
The first conversation is free. You get a rough size in hours, the tier that suits the work, and what a first 150 hour block would cover.
Or get a ballpark proposal in two minutes