How Gemini AI generates civic complaint descriptions

The core problem in CivicBridge was friction. A citizen spots a broken streetlight. To file a complaint, they need to describe it, locate it, categorise it, and phrase it in a way a municipality can act on. Most people abandon the form halfway through.

The solution: take a photo, drop a pin, and let Gemini write the complaint.

What the feature actually does

A citizen opens the app, taps "Report Issue," takes a photo of the problem, and confirms their GPS location. They hit submit. Gemini receives the image and coordinates, reverse-geocodes the location, classifies the issue type, and generates a formal complaint description ready to route to the relevant department.

The output looks like this:

Category: Road Infrastructure
Location: Near Sector 12 Market, Kanpur (26.4499° N, 80.3319° E)
Description: A significant pothole approximately 30cm in diameter has formed at the intersection adjacent to the market entrance. The damage poses a risk to two-wheelers and pedestrians. Immediate patching is recommended before the onset of monsoon season.

That's generated from a photo and a GPS pin. The citizen wrote nothing.

The prompt structure

Getting Gemini to produce consistently actionable output required a few iterations. The naive prompt — "describe this civic issue" — produced text that was sometimes too casual, sometimes too vague, and often missing the location context entirely.

The final system prompt has four explicit constraints:

The JSON output constraint was the most important addition. Free-form text responses were structurally inconsistent — sometimes a paragraph, sometimes a bulleted list, sometimes with a header. Enforcing JSON meant I could reliably parse and display each field, route by category, and store structured data rather than blobs.

Handling the coordinates

Gemini doesn't reverse-geocode natively, so I pass the human-readable address alongside the coordinates. I call the Google Maps Geocoding API first, get a formatted address string, and inject it into the prompt:

async function buildPrompt(coords, formattedAddress) {
  return `
Location: ${formattedAddress} (${coords.lat.toFixed(4)}° N, ${coords.lng.toFixed(4)}° E)

Analyse the attached image and generate a civic complaint report
following the format described in your instructions.
  `.trim();
}

This matters because Gemini's training data for Indian street-level geography is uneven. Giving it "Near Sector 12 Market, Kanpur" produces better location descriptions than asking it to derive context purely from coordinates.

The category routing

Once Gemini returns a category, the complaint routes automatically to the relevant department queue. This is where the structured output pays off:

const DEPARTMENT_MAP = {
  "Road Infrastructure": "public_works",
  "Sanitation":          "municipal_health",
  "Street Lighting":     "electricity_dept",
  "Water Supply":        "water_authority",
  "Public Property":     "parks_and_estates",
  "Other":               "general_complaints",
};

function routeComplaint(complaint) {
  const dept = DEPARTMENT_MAP[complaint.category] ?? "general_complaints";
  return assignToQueue(dept, complaint);
}

If the category were free text, this map would be unreliable — Gemini might say "Roads" one time and "Road Damage" another. Constraining the output to a fixed enum in the system prompt eliminates that variability.

What the leaderboard uses

One of CivicBridge's features is a municipality resolution leaderboard — which departments are closing complaints fastest. This only works because every complaint has a structured category and a status. Gemini's consistent categorisation is what makes the leaderboard data meaningful rather than noise.

A complaint filed as "Other" because the image was ambiguous still enters the system. A human reviewer can reclassify it before routing. The AI handles the 90% case; the system degrades gracefully for the rest.

What I'd improve

The current prompt asks Gemini to do two things at once: classify and describe. Splitting these into sequential calls — classify first, then describe with the category already confirmed — would likely improve accuracy on ambiguous images. The first call is cheap (short output); the second call benefits from the narrowed context.

I'd also add confidence signalling. Right now, a blurry photo and a sharp photo get the same treatment. Gemini's API supports asking for uncertainty in the response, and surfacing "Low confidence — please verify on site" to the citizen reduces trust erosion when the description misses.

The core insight remains: the less you ask a citizen to type, the more complaints actually get filed. Gemini doesn't replace civic engagement — it removes the activation energy barrier.

Links / Resources

• Google Gemini API documentation
• Google Maps Geocoding API