Turning raw user feedback into actionable insight

Client Industry

Technology & SaaS · consumer product team

Key Challenge

Large volumes of raw JSON feedback going unread

Core Tech Stack

FastAPI · LangChain · Text-Bison · Gemini Pro

Deployment Model

Containerized microservice on Google Kubernetes Engine

SoftmaxAI developed a production-quality feedback-analysis pipeline on Google Cloud that converts large amounts of raw user feedback into insights which teams can act upon. Comments based on ratings are synthesized into readable summaries whereas shorter and simpler entries are classified according to their sentiment — both tasks being carried out by the same service. A FastAPI application, containerized using Docker and operated on Google Kubernetes Engine, coordinates Text-Bison and Gemini Pro via LangChain and then delivers the consolidated results through a clean JSON REST API over secure HTTPS — this being done behind an HTTPS ingress, with the images stored in Artifact Registry and access to the cluster being protected by a bastion host.

• Two jobs were tangled together. The task of combining detailed comments that are based on ratings into clear summaries is different from assessing the sentiment of shorter and simpler entries — even though both types of entries ended up mixed together in the same feed.
• Raw JSON hid the signal. The feedback was in large amounts of JSON and there was no efficient method for making sense of it, so the value contained within it was being wasted.
• Manual analysis could not scale. The process of reading and sorting the feedback by hand could not keep up with the volume, so the team was always behind the stream of incoming feedback.
• Insight needed a clean interface. The results had to be in a reliable and secure format for use by other systems — a format which teams could incorporate into their own tooling, not a single-purpose report.

The service operates as a pipeline that is native to the cluster and is deployed using containers on Google Cloud, following a left-to-right data flow from raw JSON feedback, via a FastAPI service and a model-routing stage based on LangChain, to a JSON REST API that provides aggregated data over HTTPS.

• Spec before code. First we drew up and agreed on a specification — covering the processing stages, the JSON API interface, and the exact way in which the service would be deployed and secured — before actually writing the production code.
• Model routing through LangChain. Each entry is sent to the appropriate model via a single LangChain pipeline, with Text-Bison being used to summarize the comments based on ratings and Gemini Pro being used to classify the sentiment of the simpler entries.
• Containerized and cluster-native. The service uses FastAPI and Uvicorn, is containerized using Docker, and is deployed on Google Kubernetes Engine with the images coming from Artifact Registry.
• Secured, human-owned rollout. An HTTPS ingress is in front of the cluster and a bastion host protects access for administrative purposes; all the decisions regarding the pipeline and the infrastructure were reviewed by humans before they were rolled out.

• Automated comment summarization. The detailed comments, which are based on ratings, are reduced to brief and easy-to-read summaries so that nobody has to read them sentence by sentence.
• Sentiment at a glance. Shorter and simpler entries have their sentiment automatically classified, which means that the overall mood of the feedback can be seen in aggregate.
• Aggregated, actionable insight. The summaries and the sentiment are combined in a single consolidated view which shows teams what action they should take.
• A clean JSON REST API. The results are provided via a predictable JSON REST API using secure HTTPS, making it possible to integrate them into any team's current tooling.
• Production-grade and secure by design. The service is ready to scale with feedback volume since it is running on GKE behind an HTTPS ingress and uses images from Artifact Registry together with bastion-guarded access.