The founders are Kyle Dingman<\/a>, a plaintiff-side mass torts lawyer, and Stephanie Duff-O\u2019Bryan<\/a>, a commercial litigator and former Consumer Financial Protection Bureau enforcement attorney. They built DingDuff \u2014 the name is a mash-up of their surnames \u2014 for their own practices, and say they never set out to build a company.<\/p>\n They decided to share it with other legal professionals after a result that surprised them. Dingman keeps a small set of test prompts that he reruns as models and products change, and in those comparisons, he said, an ordinary Claude account connected to DingDuff was producing research memos that matched the output of the legal-specific AI products available through their firms.<\/p>\n \u201cIt turns out memos that are basically as good as Westlaw\u2019s,\u201d Dingman told me in a recent interview. \u201cWe weren\u2019t trying to build some sort of competitor to Westlaw.\u201d<\/p>\n It is a claim the founders now make publicly on their website, where they say DingDuff outperformed their firms\u2019 Westlaw AI accounts so consistently that they both eventually stopped using them.<\/p>\n Granted, that is their own assessment, not an independent benchmark. Still, it goes to a larger question the legal AI market has been debating these days, which is how much value comes from a purpose-built legal AI platform as compared to simply giving a general-purpose model such as Claude or ChatGPT direct access to reliable legal sources.<\/p>\n The tagline on DingDuff\u2019s GitHub page<\/a> emphasizes its origin story: \u201cBuilt by lawyers, for lawyers.\u201d It grew out of the mix of anxiety and fascination that swept the profession after ChatGPT\u2019s release in late 2022.<\/p>\n \u201cI got kind of obsessed with AI,\u201d Dingman said. \u201cI was like, \u2018Oh God, the robot\u2019s going to replace me.\u2019 But also I was super excited by it.\u201d<\/p>\n Founders Stephanie Duff-O\u2019Bryan and Kyle Dingman. (Composite image generated from separate headshots by ChatGPT.)<\/strong><\/p>\n<\/div>\n He read everything he could about how the models worked, bought a $10,000 computer, and spent a weekend at the home of his father, a retired software engineer, trying to figure out whether he could train specialized legal AIs.<\/p>\n The machine soon sat gathering dust as the frontier models improved faster than anything he could build. But he said the experiments taught him something about legal AI. The limiting factor was not the AI\u2019s intelligence, he concluded, but its access to data.<\/p>\n So Dingman began downloading cases from Harvard\u2019s Caselaw Access Project<\/a> and compiling them into a database, which he ran on a $35 Raspberry Pi tucked away in his closet.<\/p>\n Why the $35 computer and not the $10,000 one? \u201cIf I accidentally break it with some bad Python code, I can just buy another,\u201d Dingman explained<\/p>\n Duff-O\u2019Bryan joined the project in mid-2025, after the DOGE-era upheaval at the CFPB froze her investigations and sent her back to private practice. Neither founder is a professional coder, so both learned as they went. Duff-O\u2019Bryan said she opened her computer\u2019s terminal for the first time since a high school C++ class.<\/p>\n Their case law source changed to CourtListener<\/a>, the Free Law Project\u2019s open database of court opinions, after they reached out to its co-founder Mike Lissner. Rather than mirror CourtListener\u2019s enormous and constantly changing corpus, DingDuff accesses it through API calls. DingDuff later added its own database of federal and state statutes and rules.<\/p>\n The DingDuff team has since grown to include two developers: Dingman\u2019s brother Brady, a former naval submarine nuclear operator who curates the statutory database, and Julia Teleki<\/a>, a software developer.<\/p>\n The company has no investors, and the service is free. \u201cI have zero investors to appease,\u201d Dingman said. \u201cI am perfectly happy running this with a tip jar indefinitely.\u201d<\/p>\n The DingDuff website sums up the design philosophy in six words: A legal connector that \u201cgets out of the AI\u2019s way.\u201d<\/p>\n \u201cThe tool itself is just a pipe from the law to Claude,\u201d Dingman said. \u201cIt doesn\u2019t tell Claude how to go analyze a legal issue. Claude\u2019s actually pretty good at it out of the box.\u201d<\/p>\n \u201cThere\u2019s a lot that you want to do as an attorney that goes beyond writing a memo,\u201d he said. \u201c\u2026 What I really want to do is just interact iteratively with the AI as I\u2019m working on a memo or a brief.\u201d The model surfaces its assumptions about the law, he said, and the lawyer can challenge them, redirect the research, or ask it to pursue a competing argument.<\/p>\n Dingman is particularly interested in work that is impractical for a human researcher on an ordinary litigation budget, such as tracing back an entire line of case law so that he can better understand it and perhaps even algorithmically process all the cases. No associate has the hours for that, and it is a much different job than producing a polished one-shot memo.<\/p>\n An AI legal research product such as Westlaw does not allow that sort of interaction, he said. \u201cIt feels like they don\u2019t trust the attorney because they guide it so much.\u201d<\/p>\n He is also critical of Thomson Reuters\u2019 recently released CoCounsel MCP connector, which I have written about here<\/a>. Rather than giving Claude access to the industry\u2019s largest corpus of legal content, he said, it only lets Claude request a research report from CoCounsel\u2019s own AI.<\/p>\n \u201cIt seems to defeat the purpose of having an MCP server,\u201d he said. \u201cAn actually really useful server would give it access to the corpus directly and let Claude figure out the answer.\u201d<\/p>\n Given that CourtListener now has an MCP server of its own, which I have also written about<\/a> and discussed on my LawNext podcast<\/a>, I asked how DingDuff differs.<\/p>\n The clearest difference, Dingman said, is statutory coverage. CourtListener\u2019s connector reaches its case law collection, PACER and RECAP data, citation networks and more, but not statutes. DingDuff adds statutes, regulations and rules of procedure and evidence for the federal system and all 50 states plus the District of Columbia.<\/p>\n That layer matters more than it might sound, Dingman said. Give Claude only opinions, and it will infer what a statute says from cases quoting it. But a court may quote only the subsection relevant to its dispute, omit a critical exception, or discuss a version that has since been amended or repealed.<\/p>\n In one of Dingman\u2019s recurring test questions, the model, without access to statutes, found older cases discussing a repealed statute and treated it as current law.<\/p>\n Building the statutory database proved harder than the founders expected. Some states offer clean APIs; others required scraping every section while tracking hierarchy, cross-references and amendments over time.<\/p>\n \u201cNobody actually has a complete corpus of all the laws,\u201d Dingman observed. \u201cYou would think that somewhere there would just be some source you can go to for all of the laws, even for just a state. But there\u2019s really not.\u201d<\/p>\n Dingman credits the Supreme Court\u2019s 2020 decision in Georgia v. Public.Resource.Org<\/em><\/a>, which held that Georgia could not copyright its annotated code, for clearing the way to add statutes from states such as Vermont and Arizona, whose codes are published by Lexis and Westlaw respectively.<\/p>\n A second difference is design. Where CourtListener\u2019s tools mirror its API endpoints \u2013 a structure, Dingman said, well suited to the academic research the API was built for \u2013 DingDuff\u2019s various tools carry task-oriented names that map to the stages of legal research an attorney actually performs, such as search opinions, retrieve opinion text and browse statute hierarchy.<\/p>\n That structure, Dingman believes, cues the model\u2019s intuitions about how to conduct research without heavy prompting.<\/p>\n DingDuff\u2019s hands-off design also means the tool will sometimes get things wrong, the founders admit. Their answer is not to construct more guardrails, but to limit access to licensed attorneys, who must confirm their status in the terms of service. The guardrail is the lawyer reading and checking the output.<\/p>\n The founders say they have never seen the tool fabricate a case or statute when working from DingDuff\u2019s sources. \u201cAs long as it has access to the cases, it won\u2019t give you made-up cases,\u201d Dingman said. \u201cAll of the cases will be real, but it\u2019ll misunderstand cases.<\/p>\n \u201cThe kinds of mistakes it makes are the kinds of mistakes associates make. They\u2019re actually very human kinds of mistakes.\u201d<\/p>\n Dingman offers the real-life example of a major law firm partner who emailed him asserting that his claim was time-barred under a one-year limitations period. The partner based his email on a case quoting a statute, but the statute had since been amended.<\/p>\n \u201cI got to write them a very fun email back pointing out that you have to pull the actual statute,\u201d he said.<\/p>\n DingDuff has built that type of verification directly into the product. Alongside the connector, it offers free downloadable skills for legal research, legal analysis and citation formatting, along with what may be its most distinctive feature: a deterministic citation checker.<\/p>\n DingDuff offers four supplemental skills that enhance its performance for specific tasks.<\/strong><\/p>\n<\/div>\n After a lawyer drafts a memo or brief, the checker downloads every cited case and statute, pairs each claim in the document with the passage in the source that supports it, and opens a side-by-side review panel where the attorney marks each citation verified, in need of attention, or rejected.<\/p>\n In a demonstration during our interview, the checker tracked 79 citation instances in a draft memo. Where it verified a citation, it displayed the source text with the supporting language highlighted. Where no copy of a cited authority could be retrieved, it flagged the citation in red and told the reviewer to obtain the source and rerun the check, rather than silently approving it.<\/p>\n Because the text matching is done by code, not by the AI\u2019s memory, a made-up or misremembered quotation cannot slip through as verified. What the checker cannot do is decide whether a source actually supports the proposition, or whether the authority remains good law. That judgment stays with the lawyer.<\/p>\n Even so, Duff-O\u2019Bryan said she is still not comfortable filing anything without reading the cases herself and running a separate citator check. \u201cThe AI would just have to be bulletproof, and we\u2019re not there yet.\u201d<\/p>\n From a confidentiality standpoint, DingDuff\u2019s servers see the tool calls Claude sends them, but not the surrounding conversation, and the company\u2019s privacy policy says tool-call parameters are retained for no more than seven days.<\/p>\n The bottom line was that Claude Fable, Claude\u2019s latest model, with DingDuff scored a perfect 11\/11 on legal accuracy and had every citation correct, outperforming all the other models they tested.<\/p>\n More details on the benchmarking can be found on DingDuff\u2019s website.<\/p>\n This is somewhat of a side note, but this morning, in search of a good screenshot example, I submitted a legal question to Claude. It appeared to me that, in generating the answer, it did not use the DingDuff connector I therefore asked Claude to re-run the query, but this time to use DingDuff.<\/p>\n Notably, after running the same query through DingDuff, Claude changed its answer. Here is how Claude put it: \u201cRunning it through DingDuff\u2019s case-law tools surfaces the same authorities but sharpens the doctrine in a way worth flagging \u2014 and corrects something in my first answer.\u201d<\/p>\n Notably, by reading the actual case, Claude found that something it had first described as a common-law qualified privilege was not that \u2014 and that, in fact, the majority opinion had expressly declined to opine on the common-law rule. Rather, the opinion upheld the lower court based on a finding that the trial judge had acted within a broad measure of discretion.<\/p>\n During our interview, I had asked Dingman whether I need to tell Claude to use DingDuff or whether it would know to use it.<\/p>\n He said Claude would look at its available tools and decide which is best to answer the particular question. But if you have multiple legal research engines connected, it might try to pick one.<\/p>\n To avoid any confusion, Dingman has created a general instruction<\/a> that can be used to tell Claude to use DingDuff when it is answering a legal question.<\/p>\n His recommended instruction, which you would add to your general Claude settings, is: \u201cWhen I ask a legal question, always use the DingDuff connector and rely only on the DingDuff database. Tell me the jurisdiction you searched and include citations.\u201d<\/p>\nFrom A Raspberry Pi In A Closet<\/strong><\/h3>\n
![\"\"](\"https:\/\/i0.wp.com\/www.lawnext.com\/wp-content\/uploads\/2026\/07\/DuffOBryan-Dingman-ChatGPT-1024x768.png?resize=451%2C338&ssl=1\" "\\"\\"")
<\/a><\/p>\nGetting Out Of The AI\u2019s Way<\/strong><\/h3>\n
Adding Statutes to Case Law<\/h3>\n
The Lawyer Is The Guardrail<\/strong><\/h3>\n
![\"\"](\"https:\/\/i0.wp.com\/www.lawnext.com\/wp-content\/uploads\/2026\/07\/Screenshot-2026-07-22-095253.png?resize=500%2C446&ssl=1\" "\\"\\"")
<\/a><\/p>\nThe Benchmark: Outperforming CoCounsel and Prot\u00e9g\u00e9<\/h3>\n
![\"\"](\"https:\/\/i0.wp.com\/www.lawnext.com\/wp-content\/uploads\/2026\/07\/Screenshot-2026-07-22-102013.png?resize=450%2C379&ssl=1\" "\\"\\"")
<\/a>In a benchmark published on DingDuff\u2019s website, it gave Claude and ChatGPT, armed with the DingDuff connector, the same legal research assignment as it gave Thomson Reuters CoCounsel and Lexis Prot\u00e9g\u00e9, and then graded the resulting memos based on their substantive legal accuracy and their citation accuracy.<\/p>\nDingDuff v. Claude Alone<\/strong><\/h3>\n
Taking Back The Profession<\/strong><\/h3>\n