A foundational breakthrough in image technology.
The world's image and raster compression stack was designed for hardware that no longer exists. We rebuilt it from first principles for the silicon you already own — and the results are not incremental.
Lossless. Bit-exact across CPU architectures. Benchmarked April 27, 2026 against industry-standard codecs on the same hardware.
The Inflection Point
The world is generating data faster
than infrastructure can absorb it.
The legacy stack is breaking.
Esri LERC, JPEG 2000, and JPEG-LS were designed when CPUs were scalar and storage was local. Today's pipelines stream petabytes across cloud regions and need every cycle for inference, not codec overhead. Operators pay twice — for the storage and for the compute to move it.
No one rebuilt it for modern silicon.
Every modern CPU has 256–512-bit vector units sitting idle during decode. Every decoder copies memory it shouldn't touch. Every codec applies one strategy across heterogeneous data. Compression has been treated as a solved problem for two decades. It was never solved — only good enough for hardware that no longer exists.
The Breakthrough
RIPT.
Raster Image Predictive Tiling.
An adaptive tile-based codec that independently selects the optimal compression strategy for each region of an image from an extensible library of spatial prediction models. Faster and smaller. No tradeoff curve.
Per-tile adaptivity
Smooth elevation, sparse outliers, speckled SAR, urban edges — each tile picks the predictor that wins. One file, many strategies. The image guides the codec, not the other way around.
SIMD-native
Every predictor and every transform was designed to compile to vector instructions on day one. NEON, AVX2, AVX-512, WASM SIMD128 — same algorithm, identical bytes out. Hardware-aware, not hardware-specific.
All numeric types
14 native types — int4, int8, int16, int32, int64, their unsigned counterparts, bf16, f16, f32, f64. The first codec built for AI quantization, scientific simulation, and imagery in one pipeline.
Verified Results
Across every domain we tested,
RIPT wins on size and speed.
| Domain | Compared to | Size | Speed | Notes |
|---|---|---|---|---|
| Quantized Elevation (lossless) | Esri LERC | up to −98.9% | up to 6× faster encode | 100% / 120 files |
| Multispectral satellite (u16) | Esri LERC | −20% | 9× faster | Sentinel-2 corpus |
| Aerial NAIP (u8 4-band) | Esri LERC | −49% | 9.1× faster encode | Production NAIP |
| SAR (ICEYE u16, 3-band) | LERC + Zstd | −65% | Raw bypass mode | 1.88× vs 1.14× |
| Point cloud (i32) | Esri LERC | −48% | ~3× faster | LiDAR rasters |
| LiDAR DSM (lossy, NEON) | Esri LERC | −54% | 5.76× faster encode | Apple M3 |
Lossless mode is bit-exact: every decoded value matches the original. Lossy mode bounds error per-pixel to a user-supplied tolerance. Cross-platform reproducibility verified across NEON, AVX2, AVX-512, and scalar fallback.
Competitive Landscape
The category we created
didn't exist before us.
LERC owns ~90% of the scientific-raster mindshare. RIPT is the first codec to beat it on speed and compression, simultaneously, across every data type and domain we measured.
Defensibility
12+ patentable innovations.
Built into every byte we ship.
Provisional patent filed. Each innovation below has zero documented prior art and is independently load-bearing — re-implementing RIPT requires re-implementing all of them.
Tile-adaptive predictor selection — different region, different strategy, automatically.
Anchor-Residual Predictor (ARP) — hierarchical spatial decorrelation across scales.
NearFlat — sparse outlier encoding for nearly-constant tiles. Zero prior art.
Quadratic predictor — second-order polynomial reconstruction for smooth surfaces.
Integer-preserving lossy quantization — bounded error, bit-exact at zero tolerance.
Raw bypass mode — opt-out when entropy beats prediction (e.g. SAR speckle).
SIMD-native primitives — every predictor maps to NEON / AVX2 / AVX-512 / WASM SIMD128.
18-byte self-describing header — no external metadata, cloud-native by construction.
ByteShuffle / BitShuffle filters — entropy-coder-aware byte reordering.
8×8 tile geometry — sized for L1 cache, zero-copy data paths.
Cross-platform bit-exact guarantee — ARM-encoded data decodes identically on x86.
Profile system — 39 domain-tuned configurations, swappable per workload.
Go-to-Market
The MrSID model,
rebuilt for the cloud era.
Drop-in distribution
Several of the codecs drop into the tools the geospatial market already runs — GDAL among them — as wire-compatible replacements for their pre-built codec backends. No migration, no new format to adopt. The free decoder creates ubiquitous read access and network effects; the licensed encoder captures revenue — the playbook MrSID and ECW used to build 20-year franchises, on modern infrastructure.
Ed25519-gated SDKs
Encoder unlocked by signed offline entitlements. Six tiers — developer, commercial, enterprise, government, seat-based, transaction-based — covering individual evaluation through air-gapped federal deployments. Drop-in C, Python, and WASM bindings.
Highest-pain segments first
Satellite operators, defense ISR, and PACS vendors pay the highest per-GB and per-tile costs and have the most acute latency requirements. They sign first, validate the technology in production, and become reference accounts for the long tail.
Drop-in everywhere
Cloud Optimized GeoTIFF compatible. Drop-in LERC replacement for Esri stacks. DICOM-compatible for medical PACS. Zarr / OME-Zarr ready for scientific compute. No migration project. Toggle a flag and ship.
The Ask
A foundational technology
deserves a foundational round.
We're raising our seed to lock in technical leadership across satellite, defense, medical, and AI/ML imaging — and to convert our first production wins into a category-defining standard.