Zero Gravity Backend

From team project to production.

A Spring Boot backend that started as a team project was taken to production. The project originally had only basic CRUD APIs. Authentication, infrastructure, a deployment pipeline, and AI analysis were all added on top.

API Endpoints

Infrastructure

Sending everything would blow the budget in a month.

The goal was to build a feature that analyzes emotion records via the Gemini API. Sending a full year of records meant ~55K input tokens per request. That was not sustainable on the project budget.

Reducing data would hurt analysis quality, but dropping the feature was not an option since it was core functionality.

The answer was already in the existing data.

If sending everything was off the table, the solution was to pick the most representative record from each time period. The question was how to define "representative."

The Chart API's per-period average levels and per-bucket top reasons served as sampling criteria. For each bucket, the single record closest to "that month's average level, containing the top reason" was selected.

// Reuse Chart API average levels as the sampling baseline
ChartLevelResponse levelChart = chartService.getEmotionLevelChart(
    userId, period, startDate, timezone);
 
// Send only representative records to Gemini instead of the full set
List<EmotionRecord> representativeRecords = selectRepresentativeRecords(
    emotionRecords, levelChart, period, startDate, timezone
);

One bucket, one representative record.

The period was divided into equal time units (buckets), and the single most representative record was selected from each. Simply picking the top 12 globally could cluster them in a particular month, so splitting evenly by bucket was the better approach.

Each bucket's representative was selected by a weighted score of emotion level 60% + reason match 40%. Daily records, written as a full-day reflection, received a 1.5x weight multiplier.

private double calculateMatchScore(
    EmotionRecord record, Double targetLevel, String topReason
) {
    // 1.5x weight for Daily records
    double recordLevel = record.getEmotionId() *
        (record.getEmotionRecordType() == EmotionRecord.Type.DAILY ? 1.5 : 1.0);
    double levelScore = 1.0 - (Math.abs(recordLevel - targetLevel) / 9.0);
 
    double reasonScore = record.getEmotionReasons().contains(topReason) ? 1.0 : 0.0;
 
    return (levelScore * 0.6) + (reasonScore * 0.4);
}

Ties were broken by longest diary first, most reasons, then most recent.

Example: January bucket in a Year analysis (avg emotion level: 4.5, top reason: "Work")

AI analysis results were cached for 24 hours, and the cache was invalidated for the relevant period whenever emotion records changed.

97% payload reduction, $0.002 per request

Constraints shaped the strategy. Instead of dropping the feature, the problem was solved by reusing existing data in a new context.