https://github.com/feldera/feldera-demos
Demos showing how to build with Feldera
https://github.com/feldera/feldera-demos
ai-agents avro debezium demos feldera hopsworks kafka mysql postgresql redpanda
Last synced: 2 days ago
JSON representation
Demos showing how to build with Feldera
- Host: GitHub
- URL: https://github.com/feldera/feldera-demos
- Owner: feldera
- Created: 2026-01-09T18:33:30.000Z (6 months ago)
- Default Branch: main
- Last Pushed: 2026-06-26T06:35:55.000Z (20 days ago)
- Last Synced: 2026-06-26T07:20:43.280Z (20 days ago)
- Topics: ai-agents, avro, debezium, demos, feldera, hopsworks, kafka, mysql, postgresql, redpanda
- Language: Python
- Homepage: https://feldera.com
- Size: 13.3 MB
- Stars: 5
- Watchers: 1
- Forks: 0
- Open Issues: 2
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
Feldera Demos
## ⚡ Quick setup
Demos and examples for the [Feldera](https://feldera.com) incremental compute engine.
All demos can be run inside **Claude Code** — install it first from https://claude.ai/code.
Open this repo in Claude Code:
```bash
claude .
```
Claude handles all steps automatically: Feldera setup, pipeline loading, SQL generation, and live monitoring.
> [!NOTE]
> If you want to run demos without the assistance of `claude`. Look into the
> `README.md` of the individual demo folders.
## 🚀 Available Demos
| Demo | Folder | Claude Code Command | Interfaces with | Description |
|------|--------|---------------------|--------------|-------------|
| Agentic Fraud Detection | [agentic-fraud-detection/](agentic-fraud-detection/README.md) | `/run_fraud_demo` | AI Agent (Claude) | Claude reads a real card-skimming attack report, maps signals to a live transaction pipeline, generates SQL detection views, and runs a live fraud investigator that classifies alerts in real time. |
| Agentic Fine-Grained Access | [agentic-fine-grained-access/](agentic-fine-grained-access/README.md) | `/run_fga_demo` | AI Agent (Claude) | Claude generates detection views for rapid enumeration and hot-folder attacks in a shared file system governed by recursive group permissions. A live investigator classifies flagged users and blocks SUSPICIOUS ones by pushing `is_banned=true` — Feldera revokes their access within milliseconds. |
| Fraud Detection (Delta Lake) | [fraud-detection-delta-lake/](fraud-detection-delta-lake/) | `/run_fraud_delta_lake` | Delta Lake, S3, XGBoost | Batch fraud detection reading from Delta Lake tables on S3, with XGBoost model training. |
| Debezium + Postgres | [debezium-postgres/](debezium-postgres/) | `/run_debezium_postgres` | Postgres, Debezium, Redpanda/Kafka | CDC pipeline ingesting Postgres changes via Debezium and Redpanda/Kafka. |
| Debezium + MySQL | [debezium-mysql/](debezium-mysql/) | `/run_debezium_mysql` | MySQL, Debezium, Redpanda/Kafka | CDC pipeline ingesting MySQL changes via Debezium and Redpanda/Kafka. |
| Debezium + JDBC | [debezium-jdbc/](debezium-jdbc/) | `/run_debezium_jdbc` | Postgres, Debezium, Redpanda/Kafka | CDC pipeline sinking Feldera views to Postgres via Redpanda/Kafka and JDBC sink connectors. |
| Hopsworks Integration | [hopsworks/](hopsworks/) | `/run_hopsworks` | Hopsworks, Kafka, XGBoost | Feature pipeline integration with Hopsworks feature store and Kafka, with XGBoost model training. |
| TikTok Recommender System | [tik-tok-recommender-system/](tik-tok-recommender-system/) | `/run_tiktok` | Redpanda/Kafka | TikTok-style recommendation system using Feldera and Redpanda/Kafka. |
| IVM Benchmark | [rolling-aggregate-bench/](rolling-aggregate-bench/) | `/run_rolling_aggregate_bench` | ClickHouse, PostgreSQL | Three-engine fraud detection benchmark. Feldera IVM is O(delta) per batch; ClickHouse and PostgreSQL do full O(N) scans. At 1B rows: Feldera ~3s, ClickHouse ~226s, PostgreSQL ~8,631s. |
| Agentic Token Cost (Multi-Engine IVM) | [agentic-tokens-demo/](agentic-tokens-demo/README.md) | `/run_tokens_demo` | AI Agent (Claude), ClickHouse, PostgreSQL | The accuracy/cost counterpart to the IVM benchmark: one fraud detector at the **same thresholds** on Feldera, ClickHouse, and Postgres — only IVM signal fidelity differs, so each engine flags a different set. An LLM analyses every flagged transaction, and the demo prices that analysis across frontier models; the run reveals the gap. |
## ⚙️ Pre-requisites
For a smooth experience, the following dependencies should be setup.
#### feldera
**Feldera instance**: You can use Feldera via the [free online sandbox](https://try.feldera.com) or a local Docker setup:
| Option | How |
|--------|-----|
| Remote instance (e.g. try.feldera.com) | Add `FELDERA_HOST=https://...` and `FELDERA_API_KEY=apikey:...` to `.env` |
| Docker (no account needed) | Leave `.env` as-is — Claude pulls and starts the container, docker or podman is required to be installed |
> Note that demos which rely on docker compose scripts to setup 3rd party services (postgres, mysql, kafka/redpanda) will only work with the local docker form factor.
#### fda CLI
Some demos use the fda CLI to interact with feldera.
You can find [instructions to install it in our docs](https://docs.feldera.com/interface/cli#quick-install).
```
curl -fsSL https://feldera.com/install-fda | bash
```
#### uv
Some demos use Python scripts and the Feldera Python SDK to interact with Feldera. For this a recent installation of [uv](https://docs.astral.sh/uv/getting-started/installation/) is necessary.