Catalog
86 tools, each doing one real job, from digesting a deposition to appealing a denial (15 of them are creative tools in Decosa Studio, shown here as one card). Pick one and try it on a sample; no sign-up.
Some are in Labs: they work, but aren't ready to rely on yet. Don't see yours? Suggest a tool
Depositions, briefs, privilege, diligence and records requests.
Deposition and hearing digest
A topic-by-topic digest where every sentence links to the page:line it came from, plus where the witnesses contradict each other.
9 / 9 in both runs Planted contradictions found, on a small fictional held-out set written by the same agent as the prompts (it also flagged some traps)
Filing pre-flight
A checklist pointing at the exact words in the brief: fake or misread authorities, misquotations, bad record cites and personal identifiers left in.
89% Fake citations flagged as a problem (held-out test)
Medical chronology with page cites
A dated chronology where every line cites the file, page and spot on the page, with conflicting dates, gaps in treatment and prior conditions flagged.
189 / 198 Planted medical events found (held-out test)
Privilege review and privilege log
A privilege call for every document with the reason checked against it, plus draft log entries that describe the document without giving the advice away.
83.9% / 82.3% / 79.7% Privileged or not: overall accuracy / precision / recall (held-out test)
Privileged drafting editor
A Word redline with real tracked changes and a margin comment on each, using wording taken only from the firm's own precedents.
0.82 / 0.85 Clauses found vs expert labels (CUAD): precision / recall
Report integrity
A sentence-by-sentence check of the report against the recording, with the times to cue up, plus the key events the report leaves out.
16/16 (16/16) Planted contradictions with the recording found (found exactly)
Visit notes, appeals, coding evidence, drug and device safety.
Visit copilot
One screen for the whole visit: who said what, what is still worth asking with its source, a note checked sentence by sentence before you see it, codes and prior-auth criteria, and the work note, FMLA sections and patient instructions drafted, never signed or sent. On your own GPU for real patients.
8.4% Medical terms missed by the live transcript (lower is better)
Claim denial appeal packet
A recommendation (appeal, don't, get documents first or fix the claim), the next filing deadline, a criteria checklist with quotes, and a letter when it's supported.
36 of 40 Right recommendation on whether and how to appeal (held-out test)
HCC evidence file and RADV defence
A keep, hold or delete verdict for each submitted code, with the MEAT sentences quoted, and an evidence file for the audit folder.
189 / 200 Right keep, hold or delete verdict per code (held-out test)
No Surprises Act IDR packet and eligibility screen
An eligibility screen for Federal IDR with the dated rule and arithmetic for each check, and a draft offer brief checked against the documents.
61 of 61 Eligibility verdicts right from the documents (held-out, templated)
Medicare sales-call record
A yes, no or unclear answer to each CMS marketing question with the quote and its time, and every stated benefit checked against the Summary of Benefits.
40/40 Call questions answered right (held-out set, never used for tuning)
Promotional-claims pre-check
An annotated MLR pre-review packet: each claim with the reference passage behind it, plus off-label, comparative, fair-balance and disease-claim problems.
20 / 20 Planted claim problems caught (held-out test)
Disputes, SARs, sanctions, filings and security reviews.
Reg E dispute investigation file
An investigation file per dispute: every Reg E deadline worked out with its rule, the notice's required details quoted, and the results letter checked.
0.956 (43/45) Missing details in consumer dispute notices flagged (held-out test)
SAR narrative desk
A SAR narrative in FinCEN's structure where every sentence cites its rows and every amount, date and count is recomputed from them.
340 of 356 (95.5%) Wrong numbers planted in drafted sentences caught (held-out test)
Security questionnaire answerer
A filled questionnaire built only from your approved answers, with the source of each one, gaps sent to a person, and stale answers flagged.
0.970 (64 of 66) Filled answers that were correct (held-out test)
Filing tie-out and MD&A grounding
A workpaper that ties each MD&A figure to the table cell or calculation it came from, and flags the ones that do not match.
11 in 6,039 figures (1.8 per 1,000) Wrong flags on real, untouched 10-K MD&As (held-out test)
Model-risk evidence pack
An evidence pack showing whether an LLM you rely on still behaves as validated, from which your validators can recompute every number.
8 of 12 runs Deliberate changes to the model or prompt that raised an alert (held-out test)
Green-claims substantiation check
A table of every environmental claim in your copy: banned, needs evidence or backed, with the rule, the evidence line and a safer rewrite.
43/44 (98%) Claims given the right verdict (held-out set, first run)
Rights, consent, disclosure and honest marketing checks.
Storefront accessibility pass
A ranked fix list of accessibility problems that stop shoppers buying, each with its WCAG criterion, the element and a code fix.
122 / 127 Planted accessibility problems caught (held-out test)
Editorial-control ledger
A signed record per piece showing the AI draft, every human edit and the named editor's sign-off, with a public page readers can check.
31 of 35 (0.89) Edits that changed the meaning flagged (public research test set)
Synthetic-performer disclosure and S&P pre-flight
A pre-flight report: which performers are synthetic and how we know, an AI label and C2PA marking added when needed, and script flags with a reviewer's sign-off.
27 of 29 (93%) Planted script problems found (held-out test)
Field reports, quotes, forms, SOPs, hiring and translation.
Fill a form from your papers
Your PDF form filled (or a copy list for a form on a website), every answer next to the line of your papers it came from, and the signature and the rest left for you.
1 / 491 (0.20%) Wrong answers among all answers it wrote (PDF forms)
Walkthrough-to-quote
A draft scope and quote where each line cites a moment in the video, is marked seen, said or assumed, and is priced only from your own price sheet.
33 / 33 Planted scope items found (work, damage, hazards, access)
Field reports
A finished inspection report with a checklist, issues graded by severity and measurements, each quoting what the inspector said and when.
40/41 (98%); before 38/41 (93%) Expected issues in the final report, now vs before (synthetic set)
Live translation
Live translated captions about a second behind the speaker, with names and terms kept consistent, and a short summary at the end.
55.3 / 29.9 / 87.2 English to Spanish translation scores, chrF++ / BLEU / COMET-22 (held-out test)
Tamper-evident record
Live captions, then a speaker-labelled transcript and a summary whose every sentence cites transcript lines, in a signed record that shows any edit.
14/17 council, 9/9 interview Summary sentences backed by the transcript (synthetic sessions)
Structured oral assessment
A draft score for each rubric criterion that quotes the candidate's own words, which the examiner confirms or overrides with a reason on the record.
90.8% Draft score matches the reference score exactly (held-out test)
Check answers, test agents, audit models, triage vulnerabilities.
Grounding check
Every sentence of an AI-drafted answer checked against the source passage it relies on, with the quote, and a pass, flag or block you can put in front of Send.
0.966 Unsupported sentences caught, strict gate (it flags about 30% of sentences for a look; held-out)
Private code assistant
Qwen3.8-27B behind a standard OpenAI-style /v1 endpoint: try it hosted, with a signed receipt per call, or run it on your own GPU so nothing leaves the box. Hosted replies are capped at 2,048 tokens and don't take tool calls yet.
478 / 500 Coding benchmark, 5 core tasks, self-hosted stack (out of 500)
Verified end-to-end test runs
A browser run of your plain-language test, judged only by your assertions in code, ending in a signed certificate that fails the build on a failure.
24/24 Planted bugs and faulty test users caught (held-out test)
Endpoint auditor
Find out whether an OpenAI-style endpoint really serves the model it claims, or a smaller or more compressed one. You get a signed pass, drift or fail report you can put in the vendor file, and anyone can re-check it.
40 s typical time per task
Open-model migration check
Replay prompts you've already logged on an open model. In minutes you see whether it gives the same answers, which examples differ and why, and what 1,000 requests would cost: go, no-go or not enough data yet.
14 of 15 pairings Go/no-go verdicts that match human experts' verdicts (held-out test)
Typed-judgment API
Yes/no, one-of-N, score or label answers with a probability you can set a threshold on and a reason, from a pinned model you can re-run next quarter.
90.7% / 0.023 Yes/no questions (BoolQ): accuracy / calibration error
Stories, music, video and voice, made with consent.
Stories, songs, music videos and dubbing in your own voice, each with consent recorded and the output labelled.
Decosa Studio
Make things with your people in them: a picture book read in your own voice, a song, a music video, an animatic or a dub. Add your cast once and every pack uses it.