Work / 05 Government / Public Sector
Staff at a government body lost time to scheduling forms across several meeting platforms. We built a chat assistant that understands a request in plain language and creates the meeting on the right platform.
- Client
- Government / Public Sector
- Engagement
- 6 months
- Delivered
- 2024
ServicesApplied AI Applications
- Time to book a meeting−75%
- Meetings correct the first time78% to 94%
- Concurrent conversations500+
The brief
What they came to us with.
Setting up a meeting meant choosing a platform, filling in its form and re-entering the same details elsewhere. Each platform asked for different fields in a different shape.
The body wanted people to ask in their own words and have the meeting appear, with many people able to do so at the same time.
Constraints
- Requests in plain language, not form fields
- Meetings created on each platform in its own format
- Many people chatting at the same time
- The assistant asks when something is missing, and never guesses
Before and after
What actually changed.
The same meetings, requested in a sentence instead of a form.
What we did
How a sentence becomes a meeting.
- 01
Turn a sentence into a platform request
A language model extracts who, when, how long and which platform. It then builds the payload that platform expects, so the booking is created through its own interface and not by a workaround.
Fig. 1 One request, one payload per platformDiagram - 02
Confirm before it books
The assistant reads the request back and asks about anything unclear, such as a missing guest or an ambiguous time. Only after the person confirms does it create the meeting.
Fig. 2 A conversation that ends in a booked meetingDiagram - 03
Keep the interface fast for everyone
A React interface with Redux Toolkit keeps each conversation’s state separate. The assistant service handles many conversations at once, and each platform connector is isolated, so a slow platform never stalls the rest.
Fig. 3 The three layers of the assistantDiagram
How it ran
How the work was paced.
The connectors took longer than the model, as they usually do.
Requests and platforms
Collect real requests and map each platform’s fields.
Assistant and prompts
Extract details, ask for what is missing, confirm.
Connectors and interface
One connector per platform, chat interface in React.
Concurrency test and handover
Test with many users at once, document, train.
Outcome
What it changed.
Measured on meetings booked before and after the assistant went live.
Time to book a meeting
−75%
Before3 minAfter45 s
How measured: Average time from starting a request to a confirmed booking, sampled before and after launch.
Meetings correct the first time
78% to 94%
Before78%After94%
How measured: Share of meetings booked without a later change to time, guests or platform.
Concurrent conversations
500+
BeforeOne at a timeAfter500+
How measured: Conversations handled in parallel in the load test, within the response-time target.
Built with
ReactWeb interfaces
Redux ToolkitApplication state
PythonEverything glued together
What we would tell the next team
- Confirming before booking cost a few seconds and removed most corrections.
- Each platform’s quirks lived in its connector, so the assistant stayed simple.
- Concurrency came from queuing conversations, not from a bigger machine.
More work
Contact
Book a free 15-minute call to talk through what you are working on.
How we connect
- 01Book a short call
Choose a time that works for you on our calendar.
- 02Discovery call
A short call to understand the problem, the people and the constraints.
- 03Final scope and kickoff
We send a clear scope and cost. Once you agree, we start.
Office
393, WeWork Vijohn Tower, 3rd FloorUdhyog Vihar Phase-4, GurugramHaryana 122016
Follow