News
The Dog Days Challenge: Bring Us Your Hardest Legal Problem Before Labor Day
The dog days of summer are here. Court calendars slow down, half the office is at the coast, and that giant hairy project you've been avoiding all year is still sitting there. The 20,000 forms. The mountain of audio. The backlog nobody wants to touch.
So here's the deal. Between now and Labor Day, bring us the legal workflow problem you think is impossible, and we'll build a working solution in about an hour. Live. While you watch. Then we'll run the whole job for free.
We're calling it the Dog Days Challenge, and yes, we're serious about the hour.
Why we're doing this
Over the past few months, people have started bringing us what we internally call "big jobs." These are the projects that don't fit into any product's feature list. A few real examples:
20,000 case intake forms. A client needed to extract providers and key data from intake sheets based on specific criteria, then drop it all into the Excel template they already use for batch orders. Nobody was going to do that by hand. We pointed an agent at it with an example intake sheet and the target spreadsheet, and it just worked.
54 audio files that needed to become one deliverable. Not just transcripts. Fifty-four transcriptions, combined into a single summary document with hyperlinks back to each transcript. Transcription plus synthesis plus document assembly, in one pipeline.
Hundreds of dictation files that weren't asking for transcripts at all. The attorney dictated the task. "Draft a memo on this." "Write an email to opposing counsel." "Prepare the MSJ." The job wasn't to transcribe the audio. It was to do what the audio said. So that's what the agents did: listened, understood the instruction, and produced the actual work product.
Each time, we braced for a multi-week build. Each time, it took less than an hour.
The trick is that there's no trick
We built case.dev as open legal AI infrastructure: vaults with OCR and semantic search, audio transcription, citation verification, document formatting that produces real Word files and PDFs, and sandboxed agents that can run all of it. Then we built Linc, our open source agent harness, which gives an agent durable matter context, direct access to the documents, and the local tools to actually produce work product instead of just talking about it.
When you have those pieces, a "big job" stops being a custom software project. It becomes a prompt, a vault, and a workflow. The 20,000 intake forms problem is a loop. The 54 audio files problem is a pipeline. The dictation problem is an agent that reads instructions instead of just words.
That's the whole point of infrastructure. The hard part is already built, so the last mile takes minutes.
If you already tried this yourself, we get it
Some of you have been hacking on this stuff for a year or two. You duct-taped together a script, a vector database, and an API key, and it worked great on the demo folder. Then you pointed it at the real matter and everything fell apart.
The document was longer than the context window. The RAG pipeline retrieved the wrong five chunks and confidently answered from them. The OCR choked on the scanned exhibit from 2009. The Word doc it generated wouldn't open. You got 80% of the way there, and the last 20% turned out to be the entire job.
That's not a you problem. That's the gap between a demo and infrastructure, and it's exactly the gap we built case.dev to close. Vaults handle ingestion, OCR, and page-aware retrieval so the agent searches the full document set instead of whatever fit in the prompt. Extracted text lands on disk where the agent reads the relevant ranges instead of stuffing a 400-page deposition into a context window. Search results carry real page references. Generated Word documents get structurally validated before they ever reach you. The boring plumbing that killed your weekend project is the product.
So if you hit the wall, bring us the exact thing that broke. We got you. Watching a problem that beat you get solved in an hour is either going to be very satisfying or mildly infuriating, and honestly both are fine with us.
The other unlock: open weight models got really good
There's a second reason we can do this now, and it's worth being direct about. The latest generation of open weight models, things like Kimi K3 and DeepSeek V4, changed the economics of big jobs.
A year ago, running an agent across 20,000 intake forms meant frontier model pricing on every single document, and the math got ugly fast. Today's open weight models deliver near-frontier reasoning at a fraction of the cost per token, and because case.dev routes across 195+ models through one API, we can match each step of a workflow to the cheapest model that does it well. Extraction and classification run on efficient open models. The judgment calls that need more horsepower get routed to bigger ones. Your cost of goods drops by an order of magnitude, and so does ours, which is how a 20,000-document job becomes something we can build live on a call instead of scoping in a proposal.
The data story matters just as much, especially in legal. Open weight models can run on infrastructure you control. Through case.dev's compute layer, we can spin up dedicated GPU instances with your vault mounted as a filesystem, run the model there, and your documents never leave the environment. No third-party model provider sees your client's medical records or your privileged communications. For firms with carrier requirements, protective orders, or clients who ask hard questions about where their data goes, that's not a nice-to-have. It's the difference between "we can't use AI on this matter" and "we can."
And because our AI governance layer lets your organization control exactly which model providers are approved, you can draw the line wherever your risk tolerance sits. Some firms run everything on self-hosted open models. Some mix and match. Either way, it's your call, enforced at the infrastructure level, not a policy memo everyone hopes gets followed.
How the challenge works
- Submit your hard problem. Email us at hardproblems@casemark.com with a description of the job. The bigger and uglier, the better. Thousands of documents, messy formats, weird outputs, multi-step workflows. If it makes you tired just describing it, that's the one we want.
- We pick the gnarliest ones. Submissions close at Labor Day. We'll select entries on a rolling basis and schedule a live working session with you.
- We build it in about an hour. On a call, screen shared, clock running. You bring sample files, we bring case.dev and Linc. You watch the whole thing get built.
- We run it for free. Not just the build. The run. If your job is 20,000 documents, we process the 20,000 documents. On us.
- You keep the solution. Whatever we build in the session is yours to run again whenever you want. No strings.
What it costs you: nothing, except a story
The build is free and the run is free. In exchange, we'd like to share what we learned. High-level only: the shape of the problem, how we approached it, and roughly what it took. Something like "a firm brought us 20,000 intake forms and here's how the pipeline came together."
Nothing about you and nothing about the data unless you want the credit. Your name, your firm, and your documents stay out of it. If you'd rather be featured, great, we'll celebrate you. If you'd rather be "a mid-sized defense firm in the Pacific Northwest," that works too. Your choice, in writing, before we publish anything.
And your data is protected like it should be
We know what we're asking. You're bringing us real matter files, and in this business that can mean medical records, privileged communications, and personal information. So let's be clear about how we handle it:
- We'll enter into a contract before we touch anything. Including a Business Associate Agreement if your materials include protected health information. Signing a BAA is a normal Tuesday for us, not a special request that goes to committee.
- You get the benefit of our SOC 2 and HIPAA certified infrastructure. The same foundation our insurance carrier and law firm customers already run production workloads on. This isn't a hackathon environment. It's the real platform.
- Combined with the self-hosted open model option above, you can have your hard problem solved on certified infrastructure, under contract, with a BAA, using models where your documents never leave the environment. That's the whole stack, and you get it for a challenge entry.
The rules
- The problem has to be real. An actual job you or your firm needs done, with actual sample files we can work from. Half-finished projects that beat you count double.
- Legal work only. Litigation support, discovery, intake, transcription, drafting, research, records review. If it touches a matter, it counts.
- In about an hour means an hour of build time. If your job involves processing 20,000 documents, the processing itself runs on its own after that, and we'll cover it. We're not claiming to OCR a warehouse in 60 minutes. We're claiming the machine that does it gets built in 60 minutes.
- We get to share the high-level story, anonymized or attributed, your call. That's the only currency we're asking for.
- If we can't crack it in an hour, we'll say so publicly and tell you exactly why. We think that's a fair trade for you stumping us.
Why we're confident enough to put a clock on it
Because we've already done it, repeatedly, without meaning to turn it into a contest. The 20,000 forms, the 54 audio files, the dictation backlog. Those weren't demos. They were real client problems that turned out to be an hour of work on the right foundation.
Legal work has been priced by the hour for a century. We think it's fitting that the hour is now the unit of time it takes to eliminate a job that used to take a quarter.
Summer's almost over. That project isn't going to un-exist itself. Send it to hardproblems@casemark.com before Labor Day, then go enjoy what's left of the season while the agents do the sweating.
