This 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.

Two free ways to use us. Browse the AI Use Case Library for 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 a Structured Expert Review: we'll anonymize it, match it to experts with relevant experience, and return 3–4 perspectives with pricing.

Programming note: this usually lands on a Friday, but with the July 4th holiday weekend, I held it a few days.

Two cases this week, both from services firms that took a slow, expert-heavy process and split it into stages an agentic system could run, with people kept in control at the points that mattered. Jozef Petro of Sudolabs on wind-turbine field maintenance; Will Fernandez of Defyner on research-report production.

Key takeaways from this week's cases:

My thoughts. I interviewed Jozef on the podcast last year, and this is the build he walked me through. What I like about it is the range of technology working together: IoT sensor data off the turbines, vector search over the service manuals, and an LLM handling both the diagnosis and the on-site conversation. He runs the same stack in manufacturing, swapping in quality-inspection vision models. Different models for different jobs, wired into one workflow

The setup. A wind power plant operator was losing significant operational time to inefficient field maintenance. Technicians routinely arrived at turbine sites without the right replacement parts, and spent hours searching dense service manuals to interpret error codes once they got there. Every unnecessary return visit and every hour of manual research extended turbine downtime, costing hundreds of thousands of dollars a year in lost generation capacity.

What Sudolabs built. A custom AI mobile app for the maintenance technicians. Before each dispatch, it reads real-time IoT sensor data from the turbine flagged for service and runs an LLM with vector search across the indexed service manuals to pre-diagnose the likely failure and recommend the parts to bring. On-site, technicians describe the problem in plain language through a chat interface and get the relevant manual excerpts and repair steps back instantly. Reports write themselves after each service event. A byproduct proved as valuable as the app: the richer, more structured maintenance records became the data foundation for future predictive-maintenance modeling.

They mapped the workflow before touching the model. Sudolabs opened with a workflow analysis of the plant's maintenance operations, tracing how technicians diagnosed problems, located parts, and searched documentation until the two primary inefficiencies were clear. The app was built on LangChain and a vector database, split deliberately into a pre-dispatch diagnosis step and an on-site guidance step rather than one catch-all tool. The full engagement ran from discovery to deployment in three months, and the operator recovered its investment within roughly a year.

The value is in the orchestration. No single model does this job. The IoT sensors flag that a turbine is failing, vector search over the telemetry and the ingested manuals narrows it to the likely part, and the LLM turns dense error codes into a plain-language answer the technician can act on. The maintenance records the app captures along the way become the data foundation for predictive maintenance later. Each piece covers a gap the others leave open. Jozef runs the same stack in manufacturing, adding quality-inspection vision models alongside the LLMs to catch production defects on the line. The sensing technology and the reasoning technology do different jobs, and the data one produces feeds the next.

Best fit for: Operators at asset-intensive industrial companies (manufacturing, energy, utilities) where field-technician efficiency and equipment uptime directly drive revenue.

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

My thoughts. Here’s what "agentic" looks like when you strip away the hype: 16 narrow agents, a supervisor, and seven human review gates. The people didn't get removed. They moved to where their judgment matters.

The setup. A global research and insights business produced flagship thought-leadership reports that ran 30-plus pages, fully designed and data-rich. It shipped more than 200 a year, and each one took 24 to 32 calendar days and hundreds of active human hours to move from a structured data file to a finished, on-brand report. The work was slow, hard to scale, and tied senior people up in production instead of analysis. It also had to clear strict brand standards and several review gates at every stage, so speed could not come at the cost of quality or governance.

What Defyner built. A multi-agent report builder that turns an unstructured Excel file into a fully designed, on-brand, editable report ready for final sign-off. It runs as 16 specialist agents under a supervisor, each owning one step: interpreting the data, drafting the narrative, applying the verbal and visual brand rules, and laying out the multi-page document. Seven human review points sit along the workflow, so the team checks and steers the output at the moments that most affect quality and brand integrity. Every step is traced and logged, which makes the output auditable and its quality measurable against the agreed standard. The whole system was delivered in an eight-week sprint and runs inside the client's own environment, now in live production.

Specialist agents, not one catch-all model. The build ran as an eight-week sprint across three phases: design, develop, deploy. The first two weeks went deep on the existing report process and the brand standards a finished report had to meet, working through real past reports and mapping every step from raw file to published document. Two decisions came out of that: break the workflow into discrete stages, each handled by a specialist agent, and place seven human review points at the moments that most affected quality rather than a single check at the end. The team then built the system agent by agent on a governed, traceable foundation, testing each against real data as it came online, which took the finished cost to $14–27 per report.

Best fit for: Operators that ship high-value, structured documents at volume and to a strict brand standard (research and insights reports, investor and board decks, regulatory filings, recurring performance reporting) where the work is slow, tied up in skilled human time, and needs to keep human judgment in the loop.

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

🎙️ A strong slate of new podcast episodes is coming over the next month. In the meantime, here's one of my favorites and one of the earliest, from back when the show was called Just Curious: my conversation with Jeremy Utley of Stanford, now Harvard, on using AI as a strategic thought partner.

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 a Mini RFP brief 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.