Infrastructure
Eight GPUs, a private index and a judge model, each with its count and date
The compute, the indexes, the audits and the review process behind every figure on this site. Each number below comes from a recorded or counted run, with its date, and each says what it is not. Machines are described by class, never by name.
- Counted
GPU and CPU capacity scheduled by gpur, our job router
- 8
- GPUs4 office, 2 off-site, 2 client
- 399
- TFLOPS fp16, summedeach card's measured matrix-multiply rate
- 98 GB
- GPU memory, summedlargest single card 16 GB
- 218
- CPU threads403.8 GB memory
A sum of separate machines, not one cluster. No shard spans two cards and no memory is pooled, so the largest model a job can hold is set by the largest card. Before work is sent, each node proves itself with a real matrix multiply. Some cards sit in machines OptiVis does not own and uses with the owners' permission.
- Recorded
A private research library turned into a searchable, citable index
- 357,306
- pages extracted907 holdings, 879 works
- 1,174,245
- passage vectorsplus 344,534 page vectors
- 89 min
- to embed40 shards on 4 machines
The index is ours; the texts are not ours to redistribute. Full-text and semantic search over every page, with each passage traceable to its page. The library is private and is not published here.
- Recorded
Bulk OCR of scanned books on one rented GPU pod
- 139,513
- pages read347 scanned books
- 1
- rented A100 podterminated after the job
Burst capacity, not a standing cluster. Resumable checkpoints and per-page repair. The pod was rented for this one job and terminated when it finished.
- Recorded
Every knowledge base audited by a pinned judge model, or by a local model where text must stay on our machines
- 135,204
- judge-model judgments
- 10,409
- local-model judgments
- 23,551
- documents, deterministic checks
Model judgments, not a human review. The judge scores each passage for quality and safety; deterministic checks look for secrets, duplicates and drift. Text that must stay on our machines is judged by a local model instead. The judge's labels later trained the student classifier in reading 01.
How a figure reaches this site
Every figure on OptiVis Labs is written once in the site's source, and every page that shows it reads from there. The rules about denominators are written as tests, and the build stops if a test fails.
- RunA tool or a job produces a result, with its count and its date.
- RegistryThe figure is written once, in the site's source, with its denominator and its source.
- TestsUnit tests hold that entry to the code behind it and to the house rules.
- BuildThe site is built only if every test passes. A failure leaves the last good build live.
- PageEvery page that shows the figure reads that one entry.
Four agents from four labs, one shared review ledger
As of October 2026, OptiVis works with four coding agents from four labs: Claude, Codex, Kimi and Grok. They call each other through one router with a shared handoff ledger, a cost gate and a recursion guard. A task can be checked by one other model, and that model cannot hand it to a third.
Public copy on this site was scored by Jev, a judge model, for overclaim and for figures without a denominator before it shipped. The figures on this page were checked against their run records by a second model. Where models disagree, the disagreement is written down, not averaged away.
How a job is placed
The router cuts a job into shards small enough to lose and sends each one to a card with the memory to hold it. Turn a lane off in the diagram below to see the shards reassigned.
What this page does not claim
- No foundation model. We build on open-weight and commercial models, and we distill small task models from a judge, as in reading 01.
- No single supercomputer. Capacity is a sum of separate machines. The largest model one job can hold is set by the largest card.
- No uptime figure. These machines are counted here, and this page reports no uptime. What is verified, and how, is on the status page.
- Not every card is ours. Of the eight, two are in machines lent by their owners and two are in a client's machines, used with that client's permission. Client work runs where the written data boundary for that engagement allows.
- The library is private. The index and the method are ours. The texts are not published and are not ours to redistribute.