AI that works on your own data
RAG pipelines, MCP servers and agents, connected to the systems you already run.
We build AI that answers from your documents and your ERP, not from guesswork. It cites its sources, respects who is allowed to see what, asks before it acts, and runs on a hosted model or on your own servers.
construction.get_dpr_status(date="2026-10-20")read-only · your permissionsRetrieval-augmented generationRAG pipelines
Answers from your contracts, manuals, policies and records, with the passage they came from.
- Ingestion from file shares, email, ERPNext and other databases, kept fresh on a schedule
- Layout-aware parsing and chunking for tables, scanned PDFs and long documents
- Hybrid search: vector similarity plus keywords, then re-ranking
- Permission-aware retrieval that mirrors access in the source system
- Citations on every answer, so people can check it
- Evaluation sets that measure answer quality before and after every change
Model Context ProtocolMCP servers
MCP is the open standard that lets AI assistants use your systems as tools. We build MCP servers that give assistants safe, scoped access to ERPNext, Frappe apps and your own software.
- Tools such as check stock, create a lead or get project status
- Resources such as reports, documents and records the assistant can read
- The user's own permissions, through OAuth or API tokens, never a super-user
- Read-only by default, with an approval step before any write
- Audit log of every tool call
- Works with Claude and other MCP-compatible assistants
Multi-step work, with a person in the loop
Agents that plan, call tools and draft the result, then wait for someone to approve it.
Requirements to backlog
An RFP or a one-line idea turned into epics, user stories, use cases and non-functional requirements.
Documents to drafts
Invoices, BOQs and purchase orders read and turned into draft ERPNext records for review.
Reports that write themselves
A week of DPRs summarised into the client's weekly update, ready to edit and send.
Portals without APIs
Browser agents that fill in government and supplier portals that offer no integration.
Ask your ERP
Plain-language questions about sales, stock and projects, answered from live data.
Triage and routing
Support emails and tickets classified, summarised and sent to the right team.
Your choice of model, your rules
Hosted models
Claude, GPT and Gemini through enterprise APIs, when data policy allows.
Open-weight models
Llama, Mistral or Qwen on your own servers or a private cloud in India.
Guardrails
Scoped tools, input and output checks, and approval for actions.
Run and measure
Evaluation, monitoring and cost limits after go-live.
Built in Python and Java (including Spring AI), with LangChain or LlamaIndex where they help, and plain code where they don't.
- Pick one questionSomething your team asks every day
- Prepare the dataSources, access rules, test set
- PrototypeA working assistant on real data
- EvaluateMeasured against the test set
- HardenSecurity, approvals, monitoring
- Roll outOne team first, then the rest