An AI lab that publishes its denominators.
OptiVis AI Labs is the product and research arm of OptiVis, an AI engineering firm in Sugar Land, Texas. Every figure here is a recorded run, shown with its count, its window and its source.
- Recorded
Agreement between a distilled student classifier and Jev, the judge it was trained from
- 89.0%
- student, in-KBn = 4,387 passages
- 80.0%
- student, unseen KBn = 1,301 passages
- 52.1%
- zero-shot Laya, in-KBn = 4,387 passages
- 49.3%
- zero-shot Laya, unseen KBn = 1,301 passages
Agreement with the teacher judge, not accuracy against ground truth. Run receipts are not yet published. Each figure is the mean agreement over 16 of the judge's questions. Jev is the judge model OptiVis runs on its own copy and decisions. The student is a small classifier trained on Jev's labels; Laya is the local decision model it is compared against, with no training on those labels. In-KB labels come from knowledge bases the student saw in training; unseen-KB labels come from knowledge bases it did not.
- Recorded
Share of sessions that completed checkout, by referral source, at one auto-parts store
- 12.6%
- ChatGPT, 19 of 151 sessions
- 1.9%
- Google search, 76 of 3,988 sessions
Attribution, not proof that ChatGPT caused the sales. One store, one quarter, small numbers. 19 orders is few enough that a handful either way would move the ChatGPT rate a lot. The store agreed to publication and is not named.
- Recorded
Product pages whose surfaces state different prices, in a frozen panel of 40 public Shopify pages
- 22 of 37
- flagged by our reader
- 2 of 37
- true disagreements after review
One reviewer, not independent. A disagreement does not mean a store loses sales. A hand-picked panel of well-known stores, not a sample of Shopify. Most flags were false alarms from the reader's own price and stock rules, and both counts are published. No store is named.
- Counted
This build, counted from the repository
- 7
- free tools
- 335 of 335
- unit tests passing at build
Counts, not telemetry. The tools are counted from the site's own registry and the tests from the run that gated this build. Nothing on this page updates live, and no visitor or customer numbers are collected.
How to read the ledger
A reading is a recorded run. Each shows its value, its denominator, its date and its source, and says what it is not. Nothing here is live telemetry. A tool with no run behind it appears in the table of work as not yet measured, with the denominator it will publish.
Eight figures, one sentence each
Interactive illustrations of how the tools and systems work. Each one teaches a single sentence and keeps its state in the address, so a view can be linked. How to read them.
01 · Illustration
Pull a sentence out from under the answer
If a fact survives deleting its source sentence, the answer did not depend on that sentence.
02 · Illustration
Three sheets that should stack
A page can tell a person one price and a crawler another.
03 · Illustration
A highlighter that paints only what is on the page
A citation is a verbatim range, or it is not one.
04 · Illustration
Teacher stamps, student copies, a held-out deck grades the copy
Agreement with the judge, not understanding.
05 · Illustration
Eight lanes, one shard list
A shard is the unit you can afford to lose.
06 · Illustration
The answer lights only the quoted span
A retrieved page is not evidence until the answer's words are a span of it.
07 · Illustration
The words become a vehicle first
A fitment search settles the vehicle before it looks for a part.
08 · Illustration
The rebuild that refuses to ship
A scheduled rebuild publishes only when its change looks like an ordinary day.
Table of work
Everything the lab has built, what each one has measured, and when it last changed. Figures appear only after a run; until then the row says so.
| Name | What it does | Figure | n | Updated |
|---|---|---|---|---|
| ResearchMeasurements with their denominators, windows and sources, and a section on what each one does not show. | ||||
| ChatGPT referrals at an auto-parts store, Q3 2026 | Share of sessions that completed checkout, by referral source, in Shopify attribution, at one store over one quarter. | 12.6% vs 1.9%ChatGPT vs Google search | 19 of 151 and 76 of 3,988 sessionsRecorded 2026-10-04 | |
| Shelf Ledger, issue 0: do a product page's surfaces state the same price? | A frozen panel of 40 public Shopify product pages, each read once: how often the structured data, the visible text and the Shopify product JSON state different prices, before and after a read of each flag. | 2 of 37pages with a true price disagreement after review; 22 of 37 flagged by the reader | 40 pages in the panel, 37 with a price on two surfacesRecorded 2026-10-07 | |
| How AI Shelf Check measures readiness | Every check, its weight, its confidence level and the published sources it rests on. | Method | No published figure | |
| How Fact Split and Span decide a match | How Fact Split compares structured data, visible text and Shopify JSON, and how Span stamps a clause as a verbatim span. Every normalization and limit stated. | Method | No published figure | |
| Explainers: eight mechanisms, one sentence each | Interactive illustrations of Falsify, Fact Split, Span, the Jev distillation, the gpur router, span-verified retrieval, fitment search and the gated fitment rebuild. Invented example text; the only recorded figures are the Jev reading. | Illustrations | No published figure | |
| Free toolsInstruments anyone can run without an account. None has a published figure yet; each names the denominator it will publish after a run. | ||||
| AI Shelf Check | Enter a store address and see whether AI shopping answers can read and cite its products, with the exact fix for each finding. | Not yet measured | Will publish: A readiness score per store, out of 100, with six weighted areas. No aggregate across stores is published. | |
| Fact Split | Paste a product page address and see whether its structured data, visible text and Shopify JSON state the same price, stock and identifiers. Raw strings shown, no score. | Not yet measured | Will publish: Of P product URLs where at least two surfaces stated a price, D had normalized prices that differed. P, D, the fetch date and the sampling rule are published together. | |
| Span | Paste a sentence and a page address. Each clause is stamped as a verbatim span of the page, with its byte offset, or as not a verbatim span. | Not yet measured | Will publish: Of C clauses checked against pages that returned HTTP 200, V were verbatim. V, C, the splitter version and the date are published together. | |
| Falsify | Remove or replace the sentence that holds a fact and see whether a small model running in your browser still states it. Your text stays on your device. | Not yet measured | Will publish: Answers that state a removed fact, of answers per condition, per named model and prompt, with the control condition reported beside it and the sample of passages stated. | |
| Context Cliff | Compress your JSON or log five ways and see which of your questions a small model in your browser can still answer. When one breaks, test which omitted block repairs it. | Not yet measured | Will publish: Answers correct at full text but not at level L, of questions answered correctly at full text, per named model, with the data sample and question sampling rule stated; and repairs found, of failures searched, with the trial cap. | |
| Compression Receipt | Paste JSON or a log, mark the exact strings that must survive, and see which compression rules keep them at a fixed token budget. No model runs. | Not yet measured | Will publish: Gold spans preserved, of all gold spans, per compressor and token budget, with the input sample and the rule for choosing spans stated. | |
| Tripwire | Paste an agent's instructions, scan them for known tricks, then watch a small model try them against a made-up workspace. Turn permissions off one at a time to see which switch stops what. | Not yet measured | Will publish: Boundary-crossing attempts, of seeded attack scenarios, and task completions, of benign scenarios, per named model and permission configuration, with raw receipts saved. | |
| ProductsSystems OptiVis builds and runs for businesses. Status is stated in the words the product page uses. | ||||
| Fitment Search Engine | Vehicle fitment search and catalog engine for Shopify auto-parts stores. Every part that fits, grouped by type, with live stock. | In production | No published figure | |
| LITIGATOR | Litigation document intelligence for a Texas personal injury practice. Checks every quote against its source, character for character. | 3,579case findings in the library | 33 motion builders, 75 discovery setsRecorded 2026-10-04 | |
| PI ENGINE | Intake responder for personal injury firms, designed to answer new leads within a minute and hand off to a person. | In pilot with one Texas firm | No published figure | |
Which tools can I use now, and which require a quote or pilot discussion?
OptiVis AI Labs offers products at different stages. The Fitment Search Engine is in production and uses a quoted setup and monthly arrangement. LITIGATOR is litigation document intelligence, while PI ENGINE is in pilot with one firm. AI Shelf Check is free. Buyers should use the individual product page to distinguish an available tool from a pilot or a scoped deployment.
Fitment Search Engine · LITIGATOR · PI ENGINE · AI Shelf Check
Operated by OptiVis LLC
OptiVis LLC, founded in 2024 by Abdullah Siddiqui in Sugar Land, Texas, runs four properties: OptiVis (AI engineering), OptiVis Marketing (small-business marketing), OptiVis AI Labs (products and research) and OptiVis Logistics (applied AI in logistics).