Mixed Precision Engine

Release: Beta — suitable for trial and evaluation; APIs may change before GA

Semantic precision reduction, certified streaming, and RF detection

The Mixed Precision Engine understands numerical data at the field level — determining how many significant digits actually matter — and exploits that for compression, profiling, forensics, and optimized delivery.

Semantic field analysis, RECS v1 certified streaming compression, codec selection, and an RF FFT→matched-filter→CFAR detection pipeline ship in one package.

What it does

Enterprise data is mostly numbers stored as FP64 — often with far more precision than any downstream consumer uses. Generic compressors (zstd, LZ4) treat those bytes as opaque and achieve ~1.5×. Mixed Precision reads the semantics: for each field it asks "how many digits actually matter?" then reduces precision before compression.

The shipped library covers three workflows customers use today: semantic compression (analyze → reduce → verify roundtrip), certified streaming (RECS v1 wire frames for sensor telemetry), and the RF detection pipeline (FFT → matched filter → CFAR with per-stage precision choices).

When to use it

Use when you store or move large numerical datasets — telemetry, Parquet columns, simulation outputs, RF IQ captures — and want 3–90× compression without losing certified accuracy.

See Examples for runnable code.

Who it is for

• Data platform teams storing or moving large volumes of numerical data
• IoT operators with high-rate sensor telemetry
• RF/spectrum teams needing on-prem detection pipelines

What ships

• Python wheel (mixed_precision) with C hot paths
• Semantic compression: analyze, reduce, verify roundtrip
• RECS v1 certified streaming (mixed_precision.streaming)
• RF product pipeline (run_rf_product, run_rf_detection)