The problem
Forms put people off giving or asking for help. A chat is easier, but an AI that writes to a database is dangerous if it can be talked into acting for someone else, if it loops, or if it stops answering halfway through a donation.
AI case study · Function calling · Charity
Donating by conversation, safely
People donate goods or ask an NGO for help just by chatting. The assistant fills in the forms for them, creating and tracking real records, without ever being trusted with who the user is.
- Role
- Fullstack Developer
- Model
- Gemini, native function calling
- Timeline
- October 2026
How one message becomes a donation
- User“I have two bags of children’s clothes to give.”
- Laravel APIAdds the last 50 messages; keeps the user ID to itself.
- GeminiChooses from 10 approved tools, e.g. create_donation.
- Validated toolArguments checked; owner taken from the session.
- MySQL recordDonation saved; the reply confirms it.
How it stays safe
Security boundary
The AI never knows who you are
The backend takes the user from the login session, not from the model. Every tool argument is treated as untrusted and validated, and repositories check ownership before any read or write.
Bounded loop
Many tools, never forever
Gemini can call several tools in one turn and gets all results back together. The loop stops after 8 rounds.
Problems I hit, and how I solved them
Problem
Free-tier rate limits
Gemini’s free tier allows only a few requests a minute per model, and a single donation can take several model calls.
Solution
Wait when it’s short, switch when it isn’t
On a “too many requests” error, the backend reads how long the provider asks it to wait. If that’s a few seconds, it waits and retries once; if not, it moves straight to the next model. Everything runs inside one 25-second budget, so the user is never left hanging.
Problem
Every model is down
Out of quota, overloaded or timed out: a chat that just stops halfway through a donation loses the donor.
Solution
A guided flow takes over
A built-in step-by-step flow finishes the conversation without the AI. Its progress is remembered between messages, and if a tool already created the donation, it says so instead of creating a second one.
Problem
Context costs tokens
The model needs the conversation to make sense of “yes, those ones”, but sending all of it every time gets expensive.
Solution
Save the conversation, send a window
Every conversation is saved in the database, and only the latest 50 messages go to the model. The window is trimmed so it never starts halfway through a tool call.
Problem
Prompts stuffed with data
Loading a user’s profile and history into every prompt is slow, costly and risky.
Solution
Fetch real data only when it’s needed
The prompt holds no user data at all. Gemini calls a tool, such as list_my_donations, only when the conversation needs that fact, and gets back just that user’s records.
How it’s built
Laravel (PHP 8.3) and MySQL with thin controllers, services and repositories · React + TypeScript + Vite front end · Gemini function calling written directly against the API, no framework · six docs covering AI, API, architecture, auth, database and development.