AI agents that can take real actions, safely.
The first time an agent writes — updates a record, sends a message, moves money — is when a demo becomes a production problem.
What we build
- Tool permissions and allow-lists, so the agent can only call the actions you named.
- Shadow mode: it proposes, a person decides, until you choose otherwise.
- Human approval before any write.
- An audit log of each tool call.
- An evaluation set with regression gates in CI.
- Deployment on your infrastructure.
Proof you can open
The Policy Assistant refuses a supervisor-only document for an employee role, and create_ticket waits for supervisor approval before a ticket is created. Fourteen golden eval cases. Public source. The write-up is access control before retrieval.
How we run it
Shadow, then assisted, then supervised, then autonomous only where the evals say so. We test idempotency and crash recovery on anything that writes. Eval gates sit in CI. We use and recommend this shape; we do not claim a measured velocity multiplier.
Stack we use
The public demo runs on Gemini. OpenAI and Azure OpenAI are supported behind the same interfaces. Retrieval uses permission filters before vector search. We use durable execution (Temporal or LangGraph) when a workflow must survive a crash, and an eval harness such as a golden set in CI. The model is a choice made per engagement, not a single vendor.
The engagement
This is the In-App AI Assistant Pilot: 3–4 weeks, founding ₹4,25,000 / $13,250, list ₹8,50,000 / $26,500. The low-risk first step is the AI Readiness Audit (3–5 days, founding ₹50,000 / $1,100). There is no separate agent-sprint price.
Questions
Which models?
Gemini on the public demo. OpenAI and Azure OpenAI are supported. We pick with you, on your infrastructure.
Can it run on our infrastructure?
Yes. The pilot is deployed to your environment.
How do you test it?
A golden set of cases, including the ones that must refuse, run as a gate in CI. The public demo has 14.
What needs a person?
Anything that writes: creating a record, sending a message, moving money. Read-only answers can ship once the permission tests pass.
Who owns the code?
You do, on final payment. We keep only general tooling that predates the engagement.