A production-ready LLM observability platform built with Go and Redpanda (Kafka). StreamLens ingests telemetry events from LLM-powered applications, processes them via stream processing to compute real-time metrics, and exposes a query API for dashboards.

Portfolio Project: This project demonstrates stream processing, microservices architecture, and real-time data pipelines for LLM observability.

┌─────────────────┐
│  LLM Apps       │
└────────┬────────┘
         │ HTTP
         ▼
┌─────────────────┐      ┌──────────────┐
│ Ingestion API   │─────▶│  Redpanda    │
│  (Port 8080)    │      │  (Kafka API) │
└─────────────────┘      └──────┬───────┘
                                │
                                ▼
                    ┌───────────────────────┐
                    │ Metrics Processor     │
                    │ - Join requests/resp  │
                    │ - Window aggregation  │
                    │ - Compute metrics     │
                    └───────┬───────────────┘
                            │
                    ┌───────┴────────┐
                    ▼                ▼
            ┌──────────────┐  ┌──────────┐
            │  Postgres    │  │ Redpanda │
            │  (Metrics)   │  │ (Metrics)│
            └──────┬───────┘  └──────────┘
                   │
                   ▼
          ┌─────────────────┐
          │  Metrics API    │
          │  (Port 8081)    │
          └─────────────────┘

1. Ingestion: LLM apps send request/response events to Ingestion API
2. Streaming: Events flow through Redpanda topics (llm.requests, llm.responses)
3. Processing: Metrics Processor joins events and aggregates into 1-minute windows
4. Storage: Computed metrics stored in Postgres and published to llm.metrics topic
5. Query: Metrics API serves aggregated metrics for dashboards

• Docker & Docker Compose
• Go 1.22+ (for local development)
• Make (optional, for convenience)

# Start all services (Redpanda, Postgres, and Go services)
make docker-up

# Or manually:
docker compose -f deploy/docker-compose.yml up -d

# View logs
make docker-logs

# Wait for services to be ready (~30 seconds), then:
make generate-traffic

# Or with custom parameters:
NUM_REQUESTS=50 BASE_URL=http://localhost:8080 ./scripts/generate-traffic.sh

After ~2 minutes (to allow windowing), query the metrics:

# Get metrics for tenant-1
curl 'http://localhost:8081/v1/metrics?tenant_id=tenant-1&limit=10' | jq

# Filter by route and model
curl 'http://localhost:8081/v1/metrics?tenant_id=tenant-1&route=chat_support_v1&model=gpt-4.1-mini' | jq

1. Start infrastructure (Redpanda + Postgres):

   # In docker-compose, comment out the Go services and run:
   docker compose -f deploy/docker-compose.yml up redpanda postgres

2. Install dependencies:

   make deps

3. Run services locally (in separate terminals):

   # Terminal 1: Ingestion API
   make run-ingestion
   
   # Terminal 2: Metrics Processor
   make run-processor
   
   # Terminal 3: Metrics API
   make run-metrics-api

make build

# Binaries will be in ./bin/
./bin/ingestion-api
./bin/metrics-processor
./bin/metrics-api

Ingest an LLM request event.

Request Body:

{
  "request_id": "uuid",
  "tenant_id": "acme-corp",
  "route": "chat_support_v2",
  "model": "gpt-4.1-mini",
  "timestamp": "2025-11-19T10:00:00.000Z",
  "prompt_tokens": 321,
  "user_id_hash": "sha256-hash",
  "metadata": {
    "experiment": "prompt_v3",
    "country": "SE"
  }
}

Ingest an LLM response event.

Request Body:

{
  "request_id": "uuid",
  "timestamp": "2025-11-19T10:00:01.123Z",
  "latency_ms": 1123,
  "completion_tokens": 512,
  "finish_reason": "stop",
  "error": null
}

Query aggregated metrics.

Query Parameters:

• tenant_id (required): Filter by tenant
• route (optional): Filter by route
• model (optional): Filter by model
• limit (optional): Number of windows to return (default: 60)

Response:

{
  "metrics": [
    {
      "tenant_id": "acme-corp",
      "route": "chat_support_v2",
      "model": "gpt-4.1-mini",
      "window_start": "2025-11-19T10:00:00Z",
      "window_end": "2025-11-19T10:01:00Z",
      "requests": 1234,
      "errors": 12,
      "avg_latency_ms": 732.4,
      "p95_latency_ms": 1234.0,
      "avg_prompt_tokens": 300.1,
      "avg_completion_tokens": 420.6,
      "estimated_cost_usd": 2.31
    }
  ],
  "count": 1
}

All services are configured via environment variables:

streamlens/
├── cmd/
│   ├── ingestion-api/       # HTTP ingestion service
│   ├── metrics-processor/   # Stream processor
│   └── metrics-api/         # HTTP metrics query service
├── internal/
│   ├── config/              # Configuration management
│   ├── handlers/            # HTTP handlers
│   ├── kafka/               # Kafka producer/consumer wrappers
│   ├── models/              # Event schemas
│   ├── processor/           # Stream processing logic
│   └── store/               # Postgres storage layer
├── deploy/
│   ├── docker-compose.yml   # Docker Compose config
│   ├── Dockerfile           # Multi-service Dockerfile
│   └── init.sql             # Postgres schema
├── scripts/
│   └── generate-traffic.sh  # Sample data generator
├── Makefile                 # Development tasks
└── README.md

# Run all tests
make test

# Test specific package
go test -v ./internal/processor

• Stream Processing: Real-time joining of requests/responses by request_id
• Windowed Aggregation: 1-minute tumbling windows per (tenant, route, model)
• Metrics Computation:
  • Request count
  • Error count
  • Average & P95 latency
  • Token usage (prompt/completion)
  • Estimated cost
• Dual Sink: Metrics written to both Kafka topic and Postgres
• Graceful Shutdown: All services handle SIGTERM/SIGINT correctly
• Production-Ready: Proper error handling, logging, and connection pooling

# Ingestion API
curl http://localhost:8080/health

# Metrics API
curl http://localhost:8081/health

Access Redpanda metrics and topics at:

• Admin API: http://localhost:19644

# All services
docker compose -f deploy/docker-compose.yml logs -f

# Specific service
docker compose -f deploy/docker-compose.yml logs -f metrics-processor

# Stop services
make docker-down

# Stop and remove volumes
make clean

For a quick demo, follow the Quick Start Guide:

# Start all services
make docker-up

# Generate sample traffic
make generate-traffic

# Query metrics (wait ~2 minutes for windowing)
curl 'http://localhost:8081/v1/metrics?tenant_id=tenant-1' | jq

Sample Output:

{
  "metrics": [
    {
      "tenant_id": "tenant-1",
      "route": "chat_support_v1",
      "model": "gpt-4.1-mini",
      "window_start": "2025-11-20T10:00:00Z",
      "window_end": "2025-11-20T10:01:00Z",
      "requests": 1234,
      "errors": 12,
      "avg_latency_ms": 732.4,
      "p95_latency_ms": 1234.0,
      "avg_prompt_tokens": 300.1,
      "avg_completion_tokens": 420.6,
      "estimated_cost_usd": 2.31
    }
  ]
}

• Add llm.evaluations topic support
• Implement distributed tracing (OpenTelemetry)
• Add Prometheus metrics endpoint
• Dashboard frontend (Grafana/React)
• Multi-region deployment support
• Schema registry integration
• Rate limiting on ingestion API
• Backfilling for historical data

Contributions are welcome! Please check out our Contributing Guide to get started.

• 🐛 Report a bug
• 💡 Request a feature
• 📖 Improve documentation

Please read our Code of Conduct before contributing.

This project is licensed under the MIT License - see the LICENSE file for details.

• Redpanda - Kafka-compatible streaming platform
• franz-go - High-performance Kafka client for Go
• OpenLLMetry - OpenTelemetry for LLMs