Investment & partnership

The problem, in money

Every system that stores integers stores most of them at eight bytes per element, because the width is chosen by the type declaration and not by the data. A sensor reading between 0 and 200 needs one byte. A timestamp sampled every 60 seconds needs a base and a small index. The gap between what is declared and what is needed is paid in RAM, in disk, in memory bandwidth and in the electricity that moves it — in every database, every telemetry pipeline, every inference server and every embedded device.

And the gap is neither small nor theoretical. On a 0..200 telemetry column measured over 4 million elements, 30.52 MB becomes 3.81 MB. Against a SQLite running on the same data, SUM comes out 161× faster, disk drops 1.49× and resident memory 2.73×. Every number on this page has, in the repository, the program that prints it.

Why now, and not five years ago

Three numbers that are not ours, and that explain why this is the decade in which it starts to matter:

• 3.0× against 1.6×. Over 20 years hardware peak compute grew 3.0× every two years; DRAM bandwidth grew 1.6×, and interconnect 1.4×. Adding FLOPs stopped solving it — the side that stayed narrow is exactly the one that reading fewer bytes, with no decode step to pay, attacks head on. (IEEE Micro · arXiv:2403.14123)
• +58% to 63% in one quarter. That is how much DRAM contract prices rose in Q2 2026, in the worst shortage in nearly 15 years. An AI server uses 8 to 10 times the memory of an ordinary one. (TrendForce, April 2026)
• 945 TWh by 2030. Data-centre electricity consumption is expected to more than double this decade, with AI as the driver. Moving fewer bytes is spending less energy per query — the same arithmetic, seen from the other side. (International Energy Agency)

The commercial thesis fits in one line: the bottleneck moved, and what stayed narrow was memory and bandwidth. A format that reduces both without charging a decode step attacks the bottleneck where it is today, not where it was when the classical formats were designed.

The wedge

The field already knows how to make integers smaller: dictionaries, bit packing, delta encoding. All of them share one cost — a decode step. The bytes on disk are not the values, so something has to reconstruct them before anything can be asked of them.

Smart2Raw removes that step by construction. It classifies by range and stores in the smallest native class, so what is stored is an array the processor already knows how to read. That single decision produces a property none of the alternatives has: it cannot expand the input. Every classical format has a regime where the output is larger than what went in — dictionary encoding of a high-cardinality column measures 41.01 MB against a 30.52 MB baseline. Smart2Raw's worst case ties the baseline, because its widest class is the baseline.

The proof that already exists

The four numbers at the top of this page are counts, and anyone can produce a count. There is one piece of evidence that matters more, and it appears twice in this project's record.

The first: a defect in an already-deposited release returned truncated values with no error and a valid CRC. The project's own fuzz suite found it, it was fixed, and the defect, its minimal reproducing case and the reason twenty-five suites had missed it were all published.

The second, months later: a .s2r file built in bad faith — 64 bytes, internally consistent, with a correct CRC — made the reader allocate sixteen bytes and copy four megabytes into them. It was found by reading the loader against its own arithmetic, fixed in 3.5.1, and published with the hostile file becoming a permanent test.

Neither one had to be told to anyone. Engineering discipline is hard to assess from outside; that is what it looks like from inside.

Why it is defensible

• A dated, citable priority. Six deposits with DOIs, timestamped by a third party, describing the method in full.
• An advanced version that is not published. The open version is the proof; Premium is the product.
• A licence that converts. AGPL-3.0 means anyone building this into a closed product has to have a commercial conversation. The open version is the funnel, not the giveaway.
• A surface that is everywhere. The addressable market is not a vertical — it is every system with an integer column in it.

Where it is now, honestly

One founder, a published library at version 3.5.0, a .s2r format with measured compatibility across versions, an advanced version under development, and no external capital. What is being sought is the right conversation — capital, distribution, or a first serious deployment — not a specific cheque.

Talk

The best first message says who you are and what you would want to see. If you would rather start from the technology, everything is public: run the demonstration on your own data, read the technical scope, and reproduce the benchmarks with the command printed next to them.