Gemini API integration

AI Studio or Vertex AI

Gemini is reachable by two routes, and picking the wrong one is the most common early mistake. The Gemini Developer API, created from Google AI Studio, gives you an API key in minutes. Vertex AI wraps the same models in Google Cloud: IAM roles, regional endpoints, VPC controls and enterprise terms.

One rule we apply: never put customer data through a free tier. Terms on the free tier allow Google to use content to improve its products, while paid usage differs. Confirm the current wording before real data goes in.

What we build on Gemini

None of the projects on our work pages is built on Gemini, so we have no client example to show. The following are the builds we would propose, chosen because of what Gemini does well.

• Audio triage: voicemails and call recordings turned into transcripts, intents and next actions.
• Video and image review: product videos, site photos or inspection clips described and tagged for a catalog or a checklist.
• Long-document question answering: tender packs and manuals where the whole document fits in context, so retrieval tricks can be skipped.
• Search-grounded answers for research tools, where results must carry a link to their web sources.

For chat-style features on your own data, our custom GPT development and LLM app development pages describe the wider approach.

Worked example: voicemail triage for a service business

A home-services company gets forty voicemails a day. Someone listens to each, types a note and decides who calls back. We would replace the typing and sorting, and keep the calling human.

1. CaptureThe phone system saves each recording to storage and sends an event with the caller's number.
2. Send audio to the modelFor short clips, the file goes straight to Gemini with a prompt asking for transcript, reason for calling, urgency, address mentioned and requested time. Longer files use the Files API.
3. Constrain the outputA response schema forces fields such as urgency as one of three allowed values. Our code rejects anything outside the schema.
4. RouteEmergencies create a high-priority task and a text to the on-call technician. Routine requests land in the CRM queue in the morning.
5. ReviewEach day a manager sees ten random calls beside the AI note. Disagreements go into the next prompt revision and the test set.

A transcription error on a street name can send a technician to the wrong place, so the callback script always confirms the address aloud.

Gotchas to plan for

• Safety filters can block a response. The answer then arrives empty with a finish reason of safety. Code must treat empty text as a normal outcome, not an exception.
• Thinking tokens count toward output. A low output limit can be consumed by reasoning, leaving no visible answer. Set limits with that in mind.
• Response schemas support only a subset of JSON Schema. Complex nested rules may need simplifying, with validation in your own code.
• Preview models change or retire quickly. Production should run on stable model versions, with a re-test whenever you move.
• Grounding with Google Search has its own billing and display requirements for search suggestions. Read them before promising a UI.
• Rate limits apply per project and per model. Queue batch work and back off on errors.
• Large video files are slow to upload. Process asynchronously and show progress, instead of holding a user request open.

Timeline

The pilot runs in shadow mode first: the AI output is recorded but not acted on, so you see its mistakes before it affects a customer. The fixed quote after a short call confirms which phases your project needs. Related reading: AI integration.