About the client
The client is a music artist interested in applying AI to creative work. The goal was a conversational assistant that fans could talk to through text and audio, with responses personalised to the artist's way of speaking. What began as an experiment grew into a need for a back end that could handle real-time conversations reliably as usage grew.
Real-time, personalised conversation at low latency
- Text and audio. Convert text and audio inputs into structured data while preserving context and intent.
- High-frequency queries. Handle high-frequency queries at low latency.
- Consistent transactions. Keep transactions consistent in Firebase Realtime Database.
- Phonetic data. Retrieve and format phonetic data efficiently.
- Secured deployment. Deploy securely and scalably with infrastructure as code.
- Tuned responses. Tune Gemini API responses so they are personalised and context-aware.
What FiftyFive built
FiftyFive developed a full back-end pipeline: a modular back end on Google Cloud Run; asynchronous messaging with Pub/Sub for high-frequency queries; real-time data synchronisation with Firebase Realtime Database; a Flask service layer managing service interactions and data routing; integration of Google Gemini API responses with a third-party phonetic API; automated, repeatable deployment with Terraform.
How the solution works
Input handling
- Text and audio inputs are converted into structured data that keeps the user's context and intent, so the rest of the pipeline works with a consistent format.
Asynchronous processing
- Queries are passed through Pub/Sub, decoupling incoming requests from processing.
- Bursts of high-frequency queries queue rather than overwhelm the services, which run on Cloud Run and scale with demand.
Response generation
- A Flask service layer routes each request.
- Responses come from the Google Gemini API, tuned for personalised, context-aware replies, and are combined with phonetic data retrieved from a third-party phonetic API.
State and deployment
- Firebase Realtime Database keeps conversation data in sync in real time, with attention to transaction integrity.
- The whole stack is deployed with Terraform, so environments are reproducible and secured consistently.
Technology stack
What the solution enabled
Business transformation
Real-time conversational queries on Google Cloud
Text and audio converted into structured data
Automated, repeatable deployment
Low-latency pipeline
Modular back-end pipeline for real-time conversational queries on Google Cloud.
Personalised replies
Integration of Gemini-generated responses with phonetic metadata.
Pub/Sub and Firebase
Asynchronous processing, with real-time state.
Infrastructure as code
With Terraform, giving the client a base for further development.
FiftyFive Technologies developed the back end for a personalised conversational assistant for a European music artist who wanted fans to interact with an assistant reflecting the artist's speech patterns. The modular pipeline runs on Google Cloud Run, uses Pub/Sub for asynchronous handling of high-frequency queries, Firebase Realtime Database for real-time data sync and a Flask service layer, and combines Google Gemini API responses with phonetic data from a third-party API. Terraform automates deployment.
Technical questions
How does the conversational assistant handle high-frequency queries?
Queries are passed through Pub/Sub, decoupling incoming requests from processing. Bursts of high-frequency queries queue rather than overwhelm the services, which run on Cloud Run and scale with demand.
How are the responses personalised?
A Flask service layer routes each request. Responses come from the Google Gemini API, tuned for personalised, context-aware replies, and are combined with phonetic data retrieved from a third-party phonetic API.