AI Systems & Tech Specification

Consistent with **Dev Season of Code (DSOC) 2026** guidelines, this page discloses our integration of AI models, prompt specifications, and mathematical calculations.

AI Architectures & Prompt Schemes

Incident Visual Diagnosis

Gemini 1.5 Flash (Vision)

Triggered instantly upon photo upload. The model inspects the base64 image data and returns a structured JSON containing the detected ecological hazard class, confidence index, severity classification, and a 2-sentence localized threat summary and recommended action.

Prompt / Algorithm Spec

Analyze this image of an environmental incident... return JSON format:
{
  "detected_issue": "Short descriptive title",
  "confidence": 95, // Integer 0-100
  "severity": "Low" | "Moderate" | "High" | "Critical",
  "environmental_impact": "Detailed explanation (1-2 sentences)",
  "recommended_action": "Precise recommended mitigation/cleanup action (1-2 sentences)"
}

Before & After Recovery Monitor

Gemini 1.5 Flash (Vision)

Triggered when citizens submit progress reports with cleanup photos. The model accepts two image files (Image 1 = original hazard, Image 2 = current cleanup status) to calculate the physical recovery progress percentage, estimate pollution reduction, and log an ecosystem timeline description.

Prompt / Algorithm Spec

Compare these two images of the same location... return JSON:
{
  "improvement_pct": 75, // 0-100
  "pollution_reduced": 80, // 0-100
  "recovery_status": "Improving" | "Recovered" | "Unchanged",
  "description": "Short explanation of changes"
}

Regional Climate Synthesis

Deterministic Aggregation Engine

Generates weekly summaries based on local registry databases. Integrates SQL record frequencies, top categorizations, and coordinate hazard maps to output an overview of regional hotspots and community resolution rates.

Prompt / Algorithm Spec

Heuristic aggregation of reports table: Counts frequencies, calculates resolved ratios, extracts coordinates, and outputs formatted markdown summaries for public dashboards.

Analytical Core

5-Factor Environmental Risk Engine

To prioritize incidents, we designed a composite risk engine combining AI insights, user validation, and local density vectors:

Mathematical Model

Risk = clip(3.5*S + 1.5*V + 2.0*P + 2.0*E + 1.0*F, 0, 100)

1. AI Severity Factor (S)[2, 10]

2. Consensus Validation (V)[-5, 10]

3. Population Proximity (P)[1, 10]

4. Env Sensitivity (E)[1, 10]

5. Cluster Frequency (F)[1, 10]

*The final score is bounded between 0% and 100%, and maps directly to visual badges: Low, Moderate, High, or Critical.

Laplace Trust Model

Prevents manipulation on new incidents with low vote counts and eliminates division-by-zero risks:

Laplace Smoothing FormulaTrust = (C + 1) / (C + D + 2) * 100

Where **C** is confirms, and **D** is disputes. This guarantees an initial neutral baseline of **50%** for new submissions, preventing visual anomalies.

DSOC Summer Edition Value

Startup-Grade Scale: Bridges the gaps in municipal reporting networks by enabling automated routing, structured database auditing, and priority queue dispatching.

Predictive Remediation: Provides 3-month composed line forecasts and coordinate hotspot identification, shifting agencies from a reactive posture to a proactive posture.

Verified Outcomes: Tracks outcomes in physical metrics (e.g. kilograms of trash removed, square meters of area restored), creating transparent, measurable impact data.