Legal software

The problem

Before a firm takes a matter, it must check every party against everyone it has represented or opposed. Mid-size firms usually search their billing system letter for letter, so they miss parent companies, former names, nicknames and misspellings. Intake is a web form nobody connects to the search, and the enterprise tools are quote-only and installed by consultants.

The people involved have different needs. An associate submits the request, a conflicts analyst reviews the hits, practice leaders and general counsel approve, and finance opens the matter. Conflion gives each of seven roles its own view.

How the matching works

• Names are normalized: corporate suffixes stripped, abbreviations expanded, initials collapsed so "A.B.C. Holdings" finds "ABC Holdings".
• Three scorers combine: a phonetic key (Smith and Smyth), Jaro-Winkler for typos, and trigram overlap for spacing. Rare words weigh more than common ones.
• People match on surname first, then first name with 30+ nickname pairs (Kate and Katherine), aliases and former names.
• Strong matches expand two hops through corporate families and one hop through personal links such as spouses and officers.

Measured speed

• Name search (median): 77 ms to 1.3 ms at 50,000 parties; 535 ms to 5.3 ms at 250,000.
• Submitting an intake: 1.5 s to 37 ms at 50,000; 8.0 s to 20 ms at 250,000.
• After a colleague's change, a browser downloads 1 KB instead of reloading 200 KB.
• Server memory at 250,000 parties: 5.4 GB down to 1.8 GB.

All figures come from a benchmark script run on synthetic firms.

A model that stays on the firm's hardware

The AI runs through Ollama or any OpenAI-compatible local server, so client names never leave the building. It triages hits with a short reason, finds parties in a matter description or pasted email (any name that is not in the text is discarded), drafts a conflicts memo from the recorded decisions only, and suggests column mapping for imports. Every use is written to the audit log with the model name. The ranking model is a separate local statistical fit that uses no AI provider.

Teaching and testing inside the app

• Role-based walkthroughs (open a matter, prepare a request, approve, clear, get your firm live) that point at things but never click, submit or decide.
• A first-week checklist that ticks itself off from what the person actually does, with tests that limited roles are never shown a step for a page they can't open.
• A test environment on a fictional firm, marked by a striped banner on every page. It never writes to live data, sends email or calls an outside AI provider.

Support without sharing firm data

An administrator builds a support bundle of counts, settings switches, versions and error messages with names replaced by placeholders, reads it, then saves it. Nothing is sent automatically, and saving is written to the audit trail. To see a problem happen, support uses a test firm of made-up records at the same size and shape. The live app opens empty, with no sample data, and the desktop app ships for Windows, macOS and Linux with a first-run setup.

Taking it to market

• Research. 14+ competitors compared from public sources, and 21 gaps ranked.
• Pilot. Five firms over 90 days in four phases, measured on intake-to-cleared time, analyst hours per check and hits the old system missed.

The problem

Trademark watch services flag new filings, but someone still has to decide which hits matter, check whether the applicant is related to an existing firm client (opposing a client's subsidiary can be a conflict of interest), and turn the hit into an engagement before the 30-day opposition window closes. Those three steps usually sit in different tools and inboxes.

What it does

• Triage feed of incoming Official Gazette and §2(d) hits, ranked by score and days left, filterable by registry.
• Watched marks portfolio with classes, goods and services, registries and watch strategy.
• Clearance lab: type two marks and see the scores update live.
• Enforcement studio: from any hit, a client pitch email with flat-fee options, a cease and desist letter, notice-of-opposition grounds, and an intake file.
• Batch import of Gazette filings as CSV, JSON or XML, each scored against every watched mark.

Scoring you can explain

There is no model call. Sight is an edit-distance similarity. Sound compares phonetic consonant skeletons, so CONFLARA and KONFLARA both reduce to the same letters. Those combine into a mark score, which is blended with a goods-and-services score built from exact class matches, coordinated class groups and shared description vocabulary. Every score comes with a written rationale, which matters for legal work. The app's own labels say "AI"; this is plain algorithmic scoring.

Built to run anywhere

• Standard-library Python backend with a local database, and a single-file web app that also works straight from disk.
• A native Mac wrapper that starts the server and opens the app in its own window.
• Live trademark lookups with rate-limit tracking when an API key is present, and a built-in sample index when it isn't.

The problem

Every firm now has a list of AI ideas, from vendors, partners and its own innovation team. The list is usually ranked by enthusiasm. Ideas with no evidence look as good as proven ones, the loudest request wins, and nobody can explain afterward why one pilot was chosen over another.

What it does

• A portfolio briefing that opens with a sentence a managing partner can repeat, and rule-based insights that each show the rule behind them.
• Seven-criterion scoring with confidence ranges. A missing score is never treated as favorable, and thin evidence widens the range.
• Scenario rankings (balanced, quick wins, differentiation) that state their assumptions and show how each idea moves between them.
• Side-by-side comparison, a pilot board, a stakeholder map and editable strategy briefs that export to Markdown and PDF.

The problem

Conflicts and diligence questions are graph questions: is this fund related to our client, who controlled this company in 2023, how is this buyer connected to that target? Firms answer them from spreadsheets and vendor feeds that disagree, with no record of which source said what, and with walled matters that can leak through a search result or a count.

What it does

• A graph explorer and path finder that answers "how are these two connected" within the records the user is cleared to see.
• Evidence on every edge, with rule-based confidence that always gives its reasons, and disputed, stale or inferred links drawn differently.
• Time travel: the graph as of any date, and what began or ended between two dates.
• Ethical walls in the data layer, reversible entity resolution, saved investigations and a cited brief that exports as Markdown, PDF and JSON.

The problem

Big-firm legal AI platforms are powerful but priced and deployed for big firms, and most general chat tools give answers a lawyer cannot check. A smaller practice needs the same core moves, ask, review, draft, verify, with every answer traceable to the record and a clear line on what client data leaves the building.

What it does

• Matters that hold PDFs (with OCR), Word, Excel, PowerPoint, email and more, read and split into passages on the Mac.
• Cited answers and review tables across many documents, with each quote located in Skippy's own copy and marked verified or not.
• Drafting with redlines, exported to Word with tracked changes, and multi-step workflows such as a due diligence report or a chronology of facts.
• Cite Check, Deadlines and public research through CourtListener, the Federal Register, eCFR and SEC EDGAR.

The problem

The cheapest way to win a lawsuit is to end it early: an anti-SLAPP motion, a demurrer, a motion to compel arbitration, a statute of limitations defense. Those windows open and close in the first weeks after service, often before anyone has read the complaint closely with that question in mind. Miss one and a case that could have ended in month two runs for years.

What it does

• Ranked theories: each with the motion that carries it, the date it can be filed, the authority, what the client must supply, the other side's best answer, and how far it gets (ends the case, narrows it, or moves it).
• A response clock built from the forum's own rules, such as the 60-day anti-SLAPP window or the 30-day state response deadline, with dates exportable to Outlook.
• A claim-by-theory grid showing which theory reaches which claim, first moves for this week, and checkpoints a lawyer must confirm.
• A whole case file: PDFs, Word files and emails, read in passes when the file is large, and a draft motion outline or memo on request.

Decisions behind it

• Earliest first. Theories are ordered by when they can be filed and how much of the case they end, because the value is in the weeks saved.
• Code checks the model. Deadlines come from coded forum rules, and every quote is matched against the source text in the browser.
• A lawyer decides. Each theory lists the facts a lawyer must confirm before relying on it, such as whether the plaintiffs actually signed an arbitration agreement.
• Model choice. Opus, Sonnet or Fable, with a depth setting, so a team can trade speed against thoroughness.

Built to run where lawyers already work

It is a single page that runs inside Claude on the user's own account, so it needs no new vendor, server or install. Large case files are read in 140,000-character passes, two at a time; only facts whose quotes verify are carried forward to one final ranking pass. The real cases use public complaint text and public rulings, with sources linked in the tool.

The problem

Depositions are won on the contradictions, and finding them means holding a declaration, an email thread, a performance file and a chat log in your head at once. Associates do it by hand over days, and a missed contradiction is not recoverable once the witness has been prepared and sworn.

What it does

• Ranks contradictions as case-ending, major or useful, and flags which statements were made under oath.
• Shows both sides with the exact quote and the document it came from.
• Predicts the way out: what the witness will say to explain it away, and how to close that exit.
• Writes the deposition script as commit, confront and close questions, and lists what not to ask.
• Takes a real case file: upload documents or paste in your own.

Decisions behind it

• Quotes are checked by code. Every quoted passage is matched word for word against the source document in the browser, and anything that does not match is flagged.
• Order matters. The kill shot is one admission, the sequence to get it, and why it ends the issue.
• A lawyer decides. It prepares the examination; it does not decide what to ask.

What the recorded run showed

It read 13 documents in 256 seconds and found 9 contradictions, 3 of them case-ending, with all 32 quotes verified. In a mock deposition where Claude played a coached witness, the scripted questions produced the key admission at question 8.

The problem

Courts have begun sanctioning lawyers for citations to cases that do not exist, quotations that are not in the opinion, and holdings that say the opposite. The risk grows as more drafting is assisted by AI, and an associate cite-checking at midnight is the only barrier.

What it does

• Extracts every authority from a brief: cite, pin cite, quotations and the proposition each is cited for.
• Looks up each case in a local index built from CourtListener, and flags a citation that belongs to a different case.
• Matches every quotation against the opinion text, word for word, and marks the words that differ.
• Reads the opinion to judge whether it supports the point, and routes anything uncertain to a lawyer.
• Signs off: a table a lawyer signs, exported as a file or CSV. The sign-off states that every authority was also checked on KeyCite or Shepard’s.

Decisions behind it

• Lookup first, model second. Existence and quotation checks are plain code against real text; the model only judges support.
• No text, no pass. A case with no opinion on file is marked for a lawyer and never passed.
• Say what it did not check. Statutes and rules are listed as not checked, and even a clean result is stamped “no issues found, citator check still required.”

What the recorded run showed

On a 7-case brief with problems planted, it blocked all four: two made-up or mismatched cases, one misquote and one wrong proposition. It found no problem with the two accurate citations, which still need a citator check, and sent the rest to a lawyer rather than guessing.

The problem

Clients now send guidelines that decide what they will pay: who may bill, at what rate, how entries are written, which tasks are never billable. Finance finds violations after the invoice is rejected, weeks later, when the work cannot be corrected and the write-off is already coming.

What it does

• Extracts a rulebook from the client’s guidelines, each rule with its category, a test and the exact sentence it comes from.
• Audits every entry as payable, rewrite, reduce, not payable or confirm.
• Quotes and verifies the rule behind each finding in the page.
• Proposes narratives that meet the rule, and writes a note to the billing partner.
• Exports to CSV for the billing team.

Decisions behind it

• Before the invoice, not after. It runs on the prebill export, when the entry can still be fixed.
• No new vendor. It runs on the firm’s existing Claude and a billing export.

What the recorded run showed

It extracted 36 rules and audited the 32-entry prebill, finding that $19.3k of the $26k in fees would be cut or questioned as submitted and $7.8k of that was recoverable by rewriting or attaching approvals.

The problem

Time that is worked and not recorded is never billed. Lawyers reconstruct their day from memory at 6:40 pm and reliably remember the long tasks and forget the short ones. Across a firm, even one recovered hour a day is a very large number.

What it does

• Reads the day from the sources a lawyer already works in and accounts for every activity once.
• Drafts the entries in 0.1-hour increments with narratives written to the client’s rules.
• Runs the guard on each draft so a recovered hour does not turn into a rejected one.
• Waits for approval: nothing is saved to billing until the lawyer approves.
• Compares remembered time against found time, by matter.

Decisions behind it

• Approve first. Drafts are never posted; the lawyer approves each entry.
• Compliance in the same step. The guard’s audit is part of capture, so the narratives come out compliant.
• Say what it assumes. The $344,300 a year figure on the page is labelled simple arithmetic, not a forecast.

What the recorded run showed

On one lawyer’s day it found 7.3 hours against the 4.1 hours the lawyer would have entered, drafted 15 entries and kept $3,632 payable after the audit. Every activity is accounted for once, and no entry bills more than its logged minutes.

The problem

A conflicts search is only as good as the list of names put into it, and the list comes from an intake form with a few names. The parties that create conflicts are usually in the paperwork: a former name, a parent, an insurer, a co-defendant, a trustee.

What it does

• Reads the documents and builds a party register with each party’s role, side and relationship.
• Finds what an intake form misses: former names, parents, affiliates, insurers, counsel, and officers.
• Explains who is really adverse and why a name matters, ranked by priority.
• Quotes the source sentence for every name, checked word for word in the page.
• Exports one row per search name for the firm’s conflicts system.

Decisions behind it

• Feeds the system, does not replace it. The conflicts system stays the system of record; this fills its input.
• Every name is traceable. A name with no verified quote does not make the list.
• An analyst decides. It never clears or declines a matter.

What the recorded run showed

The litigation matter produced 28 parties against 3 on the intake form, and an M&A deal produced 33 against 3. All 61 supporting quotes were found word for word in the documents.

The problem

The first read of a new matter is where things get missed: a consent buried in a customer contract, a notice period that has already run, an amendment that quietly moved a deadline. Each practice group has its own checklist, and a rushed reviewer works from memory.

What it does

• Five workflows, each driven by a written practice checklist: deal diligence, tender and coverage, PAGA notice intake, lease abstract and dates, and BAA review against a playbook.
• Bottom line first, then the catches ranked critical, high or medium, each with a next step.
• Critical dates with the clause and the arithmetic behind each one.
• The working document for the practice: a consents tracker, a coverage chart, a PAGA issue table, a lease abstract or a clause-by-clause BAA comparison, plus a first draft such as a tender letter or litigation hold.
• What to ask the client for, and a sign-off checklist for the lawyer.

Decisions behind it

• Quotes are checked by code. Every quote is matched against the source text in the page, and one that cannot be found is flagged in red.
• Gaps are marked, not guessed. Open points are labeled for the attorney to confirm.
• A lawyer decides. Claude drafts; nothing is final until the sign-off checklist is worked through.

What the saved runs show

Each workflow opens on a saved run of a demo matter. Across the five, all 66 quotes were found in the source documents. In the deal workflow, it catches that the largest customer contract treats a change of control as an assignment needing consent, which the seller’s schedules list as needing none.

The problem

Deal lawyers check the same things by hand against a deadline: whether the schedules match the reps and the contracts, what the other side’s markup quietly changed, whether cross-references still point anywhere, and what has to be signed at closing. Each is a separate pass, often done at the last minute.

What it does

• Deal room: decisions for the deal partner and the next dates, with days to closing and to the outside date.
• Schedules: each rep set against what the schedule says and what the documents show.
• Diligence: red flags, each with a checked citation.
• Turns: every change in the seller’s markup, who it favors, and a proposed middle position.
• Integrity and Closing: cross-reference and defined-term checks, and a checklist by phase with signature pages.

Decisions behind it

• Code does what code can. The redline, cross-references, defined terms and date arithmetic are computed in the page, not by the model.
• Quotes are checked word for word against the deal documents, and anything not found is flagged.
• Grounded in the work. The page cites what deal lawyers have said in public about each task.

What the saved runs show

On the demo deal, the saved runs find 7 reps the schedules don’t match, 7 major issues in the seller’s markup and 11 of 20 closing items at risk, and the code finds 4 drafting errors. All 37 quotes in the markup analysis were found word for word in the documents.

The problem

Firms announce a win and move on. The lessons in it, such as a legal question that waited years for a motion, a notice that nearly sank a claim, or a fee award never claimed, rarely get collected, so the next team repeats the delay.

What it does

• Times the wins from the first dated event to the result.
• Finds patterns across the wins, each tied to the matters behind it.
• Asks the trial team a question for every win, rather than passing judgment.
• Four tools, each with a saved run: Day-One Test ranks issues that could end a case early; Notice Monitor checks an event log against a contract’s notice clause; Fee Recovery Builder sorts time entries for a fee motion; Win to Story turns a ruling into a release, an alert angle and a matter-experience entry.

Decisions behind it

• Questions, not findings. Each "could have gone better" point is something to ask the team.
• Arithmetic stays in code. The fee tool totals the lodestar in the page.
• One tool per lesson. Every pattern ends with something a lawyer can run on the next matter.

What the demo shows

Twelve wins produce five patterns: cases won on law that took years, contract wins that turned on notice paperwork, fee-shifting wins that need a clean fee record, wins that came only after a setback, and wins nobody wrote up for clients.

The problem

Law-firm marketing is drafted from scattered notes and checked late. The risks are specific: a "specialist" or superlative claim the rules restrict, a client named without consent, a result with no disclaimer, or a fact nobody can source.

What it does

• Attorney bio: website, short and one-line bios, representative matters and website tags.
• Client alert: the alert with a source tag on every legal statement, a LinkedIn post and a subject line.
• RFP response: a requirements matrix, easy-to-miss items, win themes and a cover letter.
• Award submission: keeps confidential matters marked for the researcher only, with a referee template.
• Model picker and revisions: Opus, Sonnet or Fable, and a box to ask for a change.

Decisions behind it

• Only facts from the source. Anything missing becomes a marked placeholder, never an invented client, result or referee.
• The rules are written in. No superlatives, no assumed pronouns, clients named only with recorded consent.
• An attorney approves. Every draft is a starting point for the responsible attorney.

What the example runs show

Each workflow opens on an example run Claude generated with the same instructions. Across the four, the review raised 18 items: 3 to fix before publishing, 8 for the attorney to verify and 7 passed checks. One fix is four required RFP items with no content yet.