Don’t compress it. Ask what it is.
AT-1 Intel is information theory as an API. Point it at any file or stream and it tells you the entropy, how compressible it is (and the storage cost), what kind of data it is, its duplicate-block ratio, and, for tables, the schema and which columns are exact functions of others. Read-only: it reads, it never reverse-engineers anyone’s format.
tell_me_about(file)
at1 intel orders.csv # orders.csv, 4.1 MB, csv, entropy 3.46 bits/byte (low, structured, compressible) # compressibility: 4.22x (~76% smaller, ~970 KB stored) · dup-blocks 0.02 # table: 200,000 rows x 5 cols # - qty: int (card 20, predictability 0.0) # derived (exact functions of other columns, store as formula, query without storing): # * total = qty*price at1 intel model.safetensors # entropy 7.99, near-random, already-compressed, ~1.0x at1 intel access.log --json # full machine-readable report
Entropy & cost
Shannon bits/byte and a real storage-cost estimate before you commit, instantly separates “already at the noise floor” from “this will shrink 10×.”
Structure & schema
Kind, duplicate-block ratio, per-column type, cardinality and predictability, and the discovered derived columns a table is hiding.
Read-only, no lock-in
A pure analysis call, nothing is stored, nothing is reverse-engineered. Meters per call; first 1,000/month free.
A different question about a file
Usually you hand a file to a compressor and get a smaller file. Here you hand over a file and get a description of it: what kind of thing it is, how much genuine information it carries, how repetitive it is, and whether it is already compressed or merely large.
That turns out to be the question people actually have before a migration. Which of these buckets is worth moving, which is already as small as it will get, and which looks like data but is in fact ten thousand copies of the same report. You can answer that for a petabyte without storing a petabyte.
It reads, it does not keep
The interface is read only. Nothing is written back, nothing is retained after the answer is returned, and the reply is a set of numbers about the file rather than any part of its contents. That is a deliberate design constraint, because the most common reason a team cannot run an assessment is that they are not permitted to send the data anywhere.