AI agents
Agents that absorb the repetitive work — drafting, triage, summarising what came in from the field, answering the same question for the thousandth time — with a person reviewing anything that reaches the public.
What this covers.
Every engagement is scoped in writing before it is built. These are the parts of it this discipline is responsible for.
Drafting assistance
First drafts of routine correspondence, notices and summaries, written against the office's own material rather than from nothing.
Request and inbox triage
Incoming messages read, categorised and routed to the person who actually handles that category.
Summarisation
Long threads, field reports and documents reduced to what a reader needs in order to decide something.
Retrieval over your own material
Answers grounded in documents the office has supplied, with the source shown next to the answer.
Human review gates
An explicit approval step wherever output leaves the organisation, built into the workflow rather than left to discipline.
Logging and evaluation
Inputs, outputs and corrections recorded, so quality can be measured over time instead of assumed.
Built for a situation, not a side.
Offices answering the same questions at a volume no one can keep up with by hand.
Teams with more documents arriving than there is reading time available.
Organisations that want AI assistance with a clear approval step rather than an autonomous system acting on their behalf.
How it is built.
Retrieval before generation
The system answers from the organisation's own documents and shows which one it used, so an answer can be checked rather than trusted.
Narrow, named tasks
An agent gets one job with a defined input and a defined output. That is what makes its behaviour reviewable at all.
Review as a property of the system
Public-facing output waits for a person by construction, not because someone remembered the policy.
Measured, then changed
Runs are logged and evaluated against real cases, so a regression is visible before it becomes a problem.
Categories of tooling, not credentials. The final choice is made per project and recorded in the scope.
What we will not build.
We build infrastructure for political work. That is not the same as building machinery for political manipulation, and the difference is written down rather than assumed.
Assistive, and attributable
These agents help the office do its own work faster. We do not build systems that impersonate a person, generate synthetic endorsements, or produce content designed to be mistaken for something it is not.
Consent first
Outreach runs on contacts who have opted in, through each channel's own approved mechanism, with an opt-out that works. We do not build around a platform's consent rules.
No profiling, no persuasion engineering
We do not build voter scoring, psychographic profiling, micro-targeted persuasion, or any system whose purpose is to change how a person votes without their knowing why they are being shown it.
A person in the loop
Anything an AI system produces for an audience outside the organisation is reviewed by a person before it is published or sent. That review is part of the software, not a rule somebody has to remember.
Nothing invented
No claim, figure, endorsement or credential goes live unless the organisation can evidence it. The rule applies to this website as strictly as it applies to yours.
Tell us what it has to do.
Describe what you are running and what this has to handle. We will tell you what we would build, what it would take, and what we would leave out.
Rajasthan first, India next.