A codec compiled for your format
Every enterprise has proprietary delimited formats no general compressor understands. AT-1 reads a sampleand synthesizes a per-column codec — integers as zig-zag delta varints, fixed-decimals as scaled integers, low-cardinality strings as dictionaries, free text verbatim — with one inference path and zero per-format code. Byte-exact, and it beats general compression because it types the columns a statistical coder can only see as bytes.
- 4.2×
- vs zlib-9 on a real telemetry format (byte-exact)
- 2.4×
- vs lzma-6 on the same — general compressors can't type columns
- 0
- lines of per-format code — one inference path handles any delimited format
- never-worse
- falls back to plain deflate/lzma when there's no structure to find
Compile, then read the codec it built
The synthesized schema is inspectable — the model the compressor built is a deliverable you can read, not a black box.
at1 codec compress telemetry.csv -o telemetry.at1cc # 984,421 -> 66,173 bytes (4.20x vs zlib, synthesized-codec) # synthesized: 5 cols, delim ',' -> int (delta-varint), # int (delta-varint), enum (dictionary), enum (dictionary), # fixed-decimal[2dp] at1 codec inspect telemetry.at1cc # read the schema at1 codec verify telemetry.at1cc # byte-exact + SHA-256
One path, any format
The same inference code handles comma, pipe, tab or semicolon files with different column types and order — proven on two unrelated formats, no per-format code.
Byte-exact + never-worse
Every container carries a SHA-256 of the original and self-checks the codec before trusting it. No structure to find? It stores plain deflate/lzma — never larger than a general compressor.
Honest refusal
If the file isn’t uniformly delimited, it says so and falls back — it never pretends to have found a structure that isn’t there.
Billed per codec compiled — first 1,000 compiles/month free. See pricing.