What it does EvaDB is a query engine that wraps AI models in SQL functions. You connect it to existing databases (PostgreSQL, SQLite, Snowflake, S3) and write standard-looking queries that invoke Hugging Face classifiers, OpenAI LLMs, YOLO object detectors, or custom PyTorch models. It handles the inference, result caching, and batching behind the scenes.

The interesting bit The project treats expensive AI calls as first-class query operators with database-style optimizations: function result caching, LLM batching to cut token costs, predicate push-down, and parallel execution across CPUs/GPUs. The README shows a query that extracts speech from video, stores it in a table, then runs ChatGPT over the transcript — all in SQL.

Key highlights

• Supports structured sources (PostgreSQL, MySQL, ClickHouse, Snowflake) and unstructured storage (local filesystem, S3)
• Pre-integrated models from Hugging Face, OpenAI, Ultralytics YOLO, plus AutoML frameworks (Ludwig, sklearn, XGBoost, statsforecast)
• Vector search via CREATE INDEX ... USING FAISS on embedding columns
• Custom model registration through CREATE FUNCTION
• Google Colab notebooks for sentiment analysis, video QA, traffic analysis, and image similarity search

Caveats

• The README is heavy on feature lists and light on performance benchmarks or latency numbers; “faster AI queries” is claimed but not quantified
• Some supported frameworks listed (e.g., “Xboost” for classification) appear to have minor naming inconsistencies
• The target of “99% of AI problems” is vague and not substantiated

Verdict Worth a look if you’re a SQL-comfortable developer who needs to sprinkle AI into existing data pipelines without orchestrating separate services. ML engineers who already manage custom training and inference stacks will likely find it constraining.