Proof of Work is the newsletter of Pluris, a network of pre-vetted applied AI experts and AI-native service organizations that specialize in helping businesses design and implement AI solutions that drive value.

Pluris is free for our clients. Two ways to use us. Browse the Pluris AI Use Case Library, hundreds of real AI implementations from our network: what got built, what it cost, and what changed once it shipped, organized around the problem you're solving. Or submit an AI need through an Expert Review: we'll anonymize it, match it to experts with relevant experience, and return 3–4 perspectives with pricing.

ClinicalNet helps patients find clinical trials they might be eligible for. That runs on a map connecting each person's condition to the trials that fit it, which is trickier than it sounds: the same condition goes by several names, and every condition nests inside broader ones. ClinicalNet was building that map by hand. It covered under 5% of the conditions they needed to serve, and doing it properly would have taken 5 to 10 full-time people. What they built instead is the lead case below.

It's one of twelve case studies PressW put into the library this week, which makes them one of its largest contributors, alongside Fractional/Ode and Eliza. The three below:

I sent Bryson Greenwood, who leads AI at PressW, five questions about how these were built, so a lot of what follows is in his words. His answer to the last one — where a PE firm or a portfolio company should point AI first — is at the bottom. He's specific about where to start and what usually isn't ready yet.

Key takeaways from this week's cases:

The setup. ClinicalNet helps patients find clinical trials they might be eligible for. Until this build, they couldn't do that without a physician in the loop. Legacy search ran on text matching that missed medical synonyms, and any replacement had to rank more than 65,000 active trials inside a five-second response budget. ClinicalNet's first attempt was to maintain the condition mapping by hand. In Bryson's words: "The team tried to maintain a very small mapping subset by hand originally, this led to less than 5% of their target domains being covered by their product as it would have taken 5-10 FTEs to do this reasonably. Automation was the only tenable solution."

What PressW built. A custom NLP engine with automated medical-condition hierarchy mapping in place of a hand-built taxonomy, custom-trained NER models, and a multi-level search architecture that returns immediate results while deeper matching runs concurrently in the background. Common condition families are pre-calculated so frequent searches short-circuit. Coverage now runs across more than a million conditions and stays current as the registry changes.

The hard part was getting the test data. Bryson is direct about where the difficulty actually sat: "building up the test data to validate whether or not the answers were good enough in a clinical setting was definitely the most operationally challenging aspect. data is hard to come by." The technical problem came second: "the most technically challenging aspect was the level of concurrency and cross checking those concurrently derived results to decide what the best combination of answers was." The 1,000+ hours the build recovered were real recurring work. "Multiple times a year the team had to conduct mass manual trial matching exercises," he says. "The team can now focus on higher value activities like origination and trial partnerships."

An idea. Find the taxonomy, mapping, or lookup table in your business that one person maintains by hand. Ask them what percentage of the domain it actually covers. ClinicalNet's answer was under 5%, and nobody had asked.

Best fit for: Operators running patient-facing or customer-facing search over a domain with a messy vocabulary, where synonym awareness and sub-second relevance both matter.

→ Want to talk to Bryson? Reply to this email and I'll make the intro. Or request one directly on the case study page.

The setup. Encore needed to catch nuanced disclosures of material non-public information inside expert-network call transcripts. The available approaches could flag risk but could not explain themselves, and a flag nobody can defend is a flag nobody acts on.

What PressW built. They refined the LLM architecture around three things: structured prompt engineering for compliance-risk detection, chain-of-thought reasoning so every decision arrives with its rationale attached, and testing benchmarked against compliance standards. Model selection was tuned to balance detection performance against processing efficiency. Encore hasn't put a number on how much accuracy improved, so this one stays qualitative.

Why they spent the budget on explainability. Bryson: "explainability is everything in an AI solution. especially in highly regulated industries, nobody trusts anything unless work is shown and that is doubly true for AI derived work. if the system is built correctly, it won't impact accuracy or latency - but doing that for every decision a model makes, in language a user understands at a glance, is as much an art as it is a science." PressW has since made it a condition of doing business: "we treat explainability as tablestakes in the solutions we develop, and in fact will no longer sell solutions without it." Encore was subsequently acquired by ACA Group, with this platform among what changed hands.

Best fit for: Compliance teams, expert-network platforms, and regulated financial-services organizations that need MNPI detection an auditor will accept.

→ Want to talk to Bryson? Reply to this email and I'll make the intro. Or request one directly on the case study page.

The setup. A prominent Austin venture capital firm was putting prospective LPs through a data room full of documents in every format. LPs could not find specific information efficiently, and the fund team absorbed the difference by answering the same questions repeatedly during diligence.

What PressW built. A chatbot fine-tuned on financial language, sitting inside the firm's existing FIS DX data room. Fund documents were pre-processed into an indexed retrieval store, every answer carries a link back to its source document, and the assistant suggests relevant follow-up questions on its own. PressW stood up answer-accuracy testing before the thing went in front of a single LP.

Building where the documents already live. Embedding inside FIS DX is what makes this replicable. No sensitive material moved, no new system had to be adopted by people outside the firm, and LPs got help at the exact moment they were reading. The fundraising team recovered 60+ hours a month, and question activity became visible, which gives the firm a read on what LPs are actually digging into.

Best fit for: Investors running a fundraise with a large, multi-format data room who want LPs to self-serve without adding load to the deal team.

→ Want to talk to Bryson? Reply to this email and I'll make the intro. Or request one directly on the case study page.

I asked Bryson where a PE firm or portfolio company should point AI first. The sponsor-side answer is up top. For portfolio companies he says it's usually one of three shapes: killing high-volume manual document processing, pulling scattered data together so people can ask it questions, or encoding a core deliverable so a ten-hour report becomes fifteen minutes of spot-checking. "Unique businesses tend to require more unique solutions," he says, "but the same principles about topline and cost centers apply."

Then the caveat, which is the more useful half. "We find that many operating companies don't have their ducks in a row re: AI readiness. They often need to move their systems to API driven design or put a data lake in place for agents to act on top of. That foundational work is often where we look first."

That tracks with all three cases above. ClinicalNet's real win was the condition-mapping layer. The data room assistant works because the documents were pre-processed into something retrievable. Encore's flags are usable because someone built the reasoning trail. In each one the interesting engineering went into the layer underneath.

💬 I wrote this week on why, in 2026, there's no excuse not to be building things. → Read it

Working on something where one of these patterns might fit? Reply to this email — I'll either make the intro to the right expert directly, or open an Expert Review and come back with 4-5 expert perspectives and pricing in a few days. Both are free, and I read every reply.

— Stu

Proof of Work: Applied AI Case Studies. The newsletter of Pluris.