Google AI Studio for Android App Prototyping

Teams often discover too late that a model can't reliably do what the feature promised: extract fields from blurry receipts, answer in Hindi with the right tone, or stay under 100 words. A few hours in AI Studio with real examples answers those questions before any Android code exists.

From prototype to app

What to prototype

• System instruction and task prompt, iterated against 20–50 real inputs, including messy ones.
• Model tier: does a fast, cheap model pass, or do you need a larger one?
• Settings: temperature and maximum output length.
• Structured output: define the JSON schema your app will parse.
• Multimodal inputs: images, audio or PDFs your feature will send.
• Safety behaviour: what happens with inappropriate or adversarial inputs.

Example: receipt extraction prototype

You extract expense data from receipt photos for an Indian expense-tracking app.
Return JSON only, matching the schema. Amounts in paise (integer). Dates in ISO-8601.
If a field is unreadable, use null. Never guess the merchant name.

{
  "type": "object",
  "properties": {
    "merchant": { "type": "string", "nullable": true },
    "totalPaise": { "type": "integer", "nullable": true },
    "date": { "type": "string", "nullable": true },
    "category": { "type": "string", "enum": ["food", "travel", "shopping", "bills", "other"] }
  },
  "required": ["category"]
}

Test with 30 real receipts (with consent, and with personal details removed): crumpled, faded, Hindi and English, and handwritten totals. Record which ones fail and why. That list becomes your evaluation set.

Moving to Android

private const val EXTRACTION_PROMPT_VERSION = "receipt-v2"   // matches what you validated in AI Studio

val model = Firebase.ai(backend = GenerativeBackend.googleAI()).generativeModel(
    modelName = "gemini-3.8-flash",                            // the tier that passed your prototype
    generationConfig = generationConfig {
        temperature = 0f
        responseMimeType = "application/json"
        responseSchema = receiptSchema                          // Kotlin version of the JSON schema above
    },
    systemInstruction = content { text(RECEIPT_SYSTEM_INSTRUCTION) },
)

Full integration (App Check, streaming and errors) is in Gemini API Integration in Android. Never copy an AI Studio API key into the app. Production calls go through Firebase AI Logic or your backend.

Prototyping tips

• Use realistic inputs from your users' world, not polished examples.
• Change one thing at a time and keep notes of prompt versions and results.
• Test failure cases deliberately: empty input, wrong language, adversarial text.
• Estimate tokens per request from the prototype to forecast cost. See Tokens and Cost.

Good candidates for prototyping

Frequently asked questions

Conclusion

Use Google AI Studio to prove feasibility with real inputs: iterate prompts, choose the model tier, define structured output and build an evaluation set. Then implement the same validated setup in Android through Firebase AI Logic, never with a raw API key.

Next step: apply AI to old code with Using AI to Understand Legacy Java Code.