ELARA·CORTEXMathematical infrastructure for complex systemsInstitute for decision systems & number theory
One engine, for everything that moves.
Re-plans a route in 0.0037 ms versus 1.1185 ms for a full recompute, a 302× reduction (TR-2026-02), online or off, on your own device. Every answer is verified: 0 infeasible outputs in 5,000 adversarial constructions. Packs data up to 80% smaller, verified byte-for-byte. Six data centres become one. Re-routes around forecast weather before it arrives.
Trial the engine on your own data at no cost. If it does not reduce your routing distance, storage footprint, or schedule cost, the subscription is refundable under the plan terms.
No account is required. Three services are available to trial: route a live trip, compress a file (verified byte-for-byte by SHA-256 round trip), or reproduce every published benchmark on the benchmark page.
One engine under everything that moves
Our routing engine reaches a lower gap to best-known than Google's OR-Tools on all six CVRPLIB-X instances we measured head-to-head (TR-2026-01). It re-plans a hundred stops in about a millisecond, and re-plans a route 302 times faster than a full recompute (TR-2026-02), online or off, and it continues navigating without a network connection, unlike connection-dependent navigation apps.
A verified engine, not orchestration around a slow solver, and not an AI that guesses: two independent proof engines, Z3 and Dafny, agree on every result with zero disagreements. Every number here is yours to test, free. Re-run any instance through OR-Tools, Gurobi or Hexaly and compare.
For the work where position has to be exact and instant: location polling up to 100 Hz, ultrafast re-routing for course correction, and guidance that keeps working in a dead zone. Built for aerospace, defence and self-driving cars.
Every byte we store and send is checked by SHA-256 round trip. Our cache-resident integrity benchmark measures 30 GB/s against the listed baselines, so nothing we keep is ever a guess.
Without optimisation you continue to carry avoidable vehicle, fuel, battery and data-centre costs. The engine is designed to reduce each of these; the measured reductions are reported above.
In development: the same engine is being extended to rocket flight stabilisation and fly-by-wire control, for aerospace, defence and autonomy.
A courier in Asia. An engineer in India. A fleet operator in the United States. One engine, for every one of them.
Most platforms hand you one map for the whole planet and leave the rest to you. We do the opposite: for every market we enter, we build a solution made for it. Local roads, the local language, pricing that fits. We are live in ten markets on four continents, and adding more.
Not on the list yet? We are coming to your country. Tell us where you operate and we will build it for you: hello@elara-cortex.com
Two things the big maps can't: out-schedule Google's own optimiser, and keep routing when the signal dies.
Plan a route, kill the signal, and it keeps navigating. Connection-dependent navigation apps require a live network link; this engine continues to re-route offline. See the dead-zone routing note for the method. Across 180,474 towns and cities, not just the largest.
On the public VRP benchmarks we ran, our scheduler beats Google's own OR-Tools, 9 out of 9 on CVRPLIB-X, and holds the time windows OR-Tools drops. Then it re-routes mid-run the moment traffic, a closure or a new pickup changes the plan, even off-signal. Track every asset, white-label the tracker. One flat-rate API, bakkie, truck or drone. Fewer kilometres, fewer litres, fewer hours.
Storage drops up to 80%, a third smaller than gzip, every byte SHA-256 verified. The 2.2 GB/s figure is a native chunker benchmark on the measured host path; the live browser upload metre can be lower because it includes upload, browser and server-demo overhead. Integrity is checked by SHA-256 on every round trip. Shrink a file free on this page, then download the free 14-day binary and run it on everything. ElaraShrink · ElaraShrink Box · ElaraShrink Server.
The Elara family
Six product lines, ten editions, one engine, one mathematics. Trial it against the metric that matters to you.
Elara V5 Pro. A 2,000,000-token mind, defined in mathematics, not just trained, so it can run its own code and check its answer before it replies. Relevance-ranked, with room for your whole problem.
Coding with Elara
One coding engine, three tiers. Reach for the fast one on everyday edits, the balanced one to build a whole feature, or the deep one for the problems that make you stop and think. Each tier reads your whole project and checks its own work before it answers.
The whole map, all 180,474 towns, runs on one small machine, not a data centre. Because the mathematics is solved, not guessed, the same engine shrinks structured data up to 80%, beating Brotli and zstd. The compression isn't a side product: it's the reason six data centres become one, and the reason everything else fits.
Our approach is deterministic and mathematically specified rather than purely learned, which makes each result reproducible and checkable.
Every claim here is free to test, right now. Drop a file and watch it beat Brotli and zstd. Run the public benchmarks against Google's OR-Tools scheduler. Ask our 2,000,000-token assistant that holds the whole thread in context. The mathematics behind it has a public, verifiable track record on the GIMPS registry. Check it all yourself.
| What we say | How it's measured | Check it yourself |
|---|---|---|
| 80% smaller than raw, a third smaller than gzip | SHA-256 round-trip: the same bytes come back, every time | Drop a file → |
| 2.2 GB/s data chunking | We rebuilt the chunker in Rust, head to head, identical boundaries | The head-to-head → |
| 30 GB/s integrity hashing, 60× faster than SHA-256 | Cache-resident, measured against xxHash and SHA-256 on the same data | The head-to-head → |
| Routes the planet from one small server | 180,474 towns on a live API, tested against the big maps | Search any town → |
| Minimised drift at 2,000,000 tokens | 100% exact recall on our published needle-in-haystack stress test; the strongest 1M-context system we tested in the same suite scored 44% | See the methodology → |
| Efficiency work with a public track record | The GIMPS public registry, of 250,000. Anyone can verify it | Look us up → |
From the team that supplied Sasol, Anglo American, Owens & Minor and Würth. ISO 9001 and 13485, two Presidential awards, 39 patents and trademarks. Meet the people →
