https://docs.google.com/document/d/1SjuPJjWqhkRerleD6uS5C-xPenZ2KEvskF1KETzcRWM/mobilebasic?tab=t.wqsoiymu8yr8 ===== Meeting - 2026/10/16 10:03 CDT - Notes by Gemini Oct 16, 2026 Meeting Invited Joe Nayigiziki Keegan Faudree Reid Valentine Attachments Meeting Meeting records Transcript  Summary Experiments, hardware requirements, and meeting applications evaluated for operational efficiency. Experiments and AI Coworker Baseline control experiments established no artificial intelligence as the standard. The primary artificial intelligence coworker was named Cletus. Mail Scanning and Architecture Bulk scanning reduced processing time for 49 documents to 1 hour. Systems utilized application programming interfaces for document movement. County Data and Applications Acquisition costs for county data repositories varied significantly. A meeting application successfully organized discussions into projects and decisions. Decisions Aligned Designation of the primary AI coworker The participants agreed to name the primary AI coworker Cletus. Document system API integration requirement The team established a non-negotiable requirement that chosen document systems must allow seamless API-based movement of files. Hardware capacity buffer sizing standard The participants agreed to size hardware capacity with a 25% buffer to accommodate future payload demands. Single region focus strategy The team aligned to focus on one region at a time instead of multiple counties simultaneously. Full documentation of meeting discussions Participants agreed to document everything said in meetings, including ideas that are not immediately used. Cletus AI coworker development The group aligned to develop the Cletus AI coworker for document handling and task execution as the primary experiment. Next steps [Joe Nayigiziki] Update link access: Change the meeting artifact privacy settings to public to allow all participants access. [The group] Confirm experiment components: Review the test artifacts and ensure all golden data sets are correctly in place for the upcoming experiment. [Joe Nayigiziki] Push code fixes: Apply the necessary fixes once the reviews are complete and push the code to the repository. [Joe Nayigiziki] Define product metrics: Draft the specific metrics for evaluating the performance and behavior of the AI coworker. [Keegan Faudree] Acquire County Records: Procure and provide Ector County records by Monday. [Joe Nayigiziki] Review Documents: Analyze research documents about Cletus system to identify agent architecture. [Reid Valentine] Develop Verification Method: Develop a programmatic method to measure verification for the experiment. [Reid Valentine] Send Results Writeup: Compile test results and methodology into a written summary for review. [The group] Deploy Application: Deploy the meet notes application to a live server. [The group] Centralize Artifacts: Establish a centralized storage location for application artifacts to ensure accessibility instead of relying on email attachments. [The group] Stream Terminal View: Integrate the call code terminal view into the application to allow users to view ongoing conversations. [Reid Valentine] Email Benchmark Results: Send the results of the conflict checks and benchmarks via email once the testing process finishes. Details Meeting Agenda and Opening Discussion: Joe Nayigiziki opened the meeting and outlined the primary topics for discussion: ongoing experiments, hardware requirements, and a new meeting site and application. After resolving initial connection issues, Joe Nayigiziki, Reid Valentine, and Keegan Faudree prepared to review progress despite a hard stop at 11:15 AM for Joe Nayigiziki (00:00:57) (00:15:44). Source Conflict Audit: Reid Valentine reported on a precursory test running a source conflict audit that compared goldens across 40 or 50 deals transferred from HubSpot against Excel inventory sheets to determine what to score work against (00:01:58). Joe Nayigiziki noted that the experimentation approach would focus on practical answers rather than overly complex scientific rigor (00:05:07). Baseline Control Experiments: Reid Valentine described a second precursory test where Claude evaluated transcripts and artifacts to determine that the control baseline should be no AI, using plain code matching on check stubs and division orders to find the minimum viable way to match documents (00:03:35). Reid Valentine shared markdown files containing practitioner stories, case studies, and research for Joe Nayigiziki to review (00:10:59). Naming the AI Coworker: Reid Valentine and Joe Nayigiziki discussed the conceptual shift of the underlying automated system toward an autonomous coworker, establishing that a human-AI replacement should have a distinct identity (00:07:08). Reid Valentine proposed the name "Cletus" for the primary AI coworker, and Joe Nayigiziki agreed with the selection (00:07:55). Physical Mail Scanning and Categorization Workflow: Keegan Faudree shared an update on training Kelsey to process physical mail packets—such as 49 property tax documents or mixed division orders and bank statements—using a bulk scan method with white paper page separators instead of scanning each page individually (00:18:30). By utilizing Claude to categorize and name the resulting PDFs, processing time for 49 documents was reduced from three and a half or four hours down to one hour, with subsequent batches taking only 15 minutes (00:23:08). Document Management and System Architecture: Keegan Faudree, Reid Valentine, and Joe Nayigiziki debated document storage systems like Dropbox, Box, and SharePoint, emphasizing that the underlying system must allow seamless movement of documents via APIs without excessive technical complexity or brittle bridges that require constant IT maintenance (00:25:18). Reid Valentine noted that they currently use a Microsoft Graph API to move documents within SharePoint via Claude and automations (00:27:24). Meeting App Demonstration and Features: Joe Nayigiziki demonstrated a newly developed meeting site and application that connects to Google Meet and Claude to listen to conversations, generate meeting notes, and automatically organize discussions into projects, topics, decisions, and to-dos on a status board (00:30:42) (00:50:21). Joe Nayigiziki tested the live Google Meet skill integration with Claude during the call to confirm functionality (00:55:29). Compound Product Development and Team Onboarding: Joe Nayigiziki, Reid Valentine, and Keegan Faudree discussed how the new meeting application acts as an AI project manager—fostering compound engineering—enabling team members like James or future hires to easily generate work plans, interact with the system without needing local Claude setups, and onboard new personnel efficiently (00:33:35) (00:39:02). Hardware Compute Buffers: Joe Nayigiziki and Keegan Faudree discussed hardware purchasing budgets and compute capacity, agreeing that systems should maintain a buffer of 25% above current forecasted payload demands to avoid future scaling limitations (00:41:51). County Data Repositories and Title Workflow: Keegan Faudree and Reid Valentine addressed acquiring county data repositories, noting that Ector County cost $2,500 while New Mexico counties like Lea County quoted excessive prices of $250,000, leading to a preference for targeted local focus and OCR text processing rather than full document deconstruction for title work (00:44:15). Keegan Faudree committed to bringing the Ector County data on Monday (00:48:50). Meeting Notes App Testing and Transcript Processing: Joe Nayigiziki tested the meeting notes application by feeding the Google Meet transcript into Claude to summarize the session (00:58:31). The application automatically parsed the transcript into decisions and to-do items. Reid Valentine and Joe discussed the application's user interface, automatic semantic categorization, and touched upon overarching project topics including scanner and storage, AI server hardware, and the development of the AI coworker named Cletus (01:02:02). Review of Automatically Categorized To-Dos and Decisions: Joe Nayigiziki and Reid Valentine reviewed the application's output, noting to-dos such as Keegan bringing Ector County records on Monday, and Joe, Reid, and Keegan running early conflict checks and setting experiment benchmarks. The participants confirmed key decisions, including focusing on one region at a time rather than multiple counties simultaneously, and building the AI coworker named Cletus to handle document wrapping, title running, and task handling. Joe requested that Claude maintain a 100% audit trail by referencing the source data for any added to-dos or decisions (01:03:05) (01:05:10). App Utility and Meeting Documentation Goals: Reid Valentine evaluated the meeting application, arguing that the tool helps structure discussions, minimize tangents, and prevent valuable ideas from falling through the cracks, a problem Reid described as data leakage (01:06:13). Reid emphasized that because the team intended to overhaul business operations, everything discussed should be documented to surface past ideas when relevant (01:07:18). Joe Nayigiziki instructed Claude to recap the discussion and questioned how Claude policed meeting goals in Claude's role as a technical program manager (01:08:08). Weekend Research Plans and Verification Benchmarks: Joe Nayigiziki stated they planned to finish research over the weekend to select the final agent systems and begin building out experimentation the following week (01:09:01). Joe assigned Reid Valentine the task of building the verification piece to programmatically measure results. Reid committed to finishing two current tests, writing up an explanation of their methodology and results, and pushing the work to a repository (01:10:02). County Records, Handwritten Indexes, and OCR Pipeline: Keegan joined the meeting and discussed acquiring Ector County, Midland County, and Martin County documents, noting the challenge that older records rely on handwritten indexes. Keegan explained that while they previously used Claude vision to match handwritten records, it excessively consumed tokens. Reid Valentine described the current OCR and vision escalation pipeline, which uses regular OCR and escalates to Gemini and Claude when page quality is low, though it still incurs token costs until a local server is established (01:11:59). The team agreed to review the document formatting when the documents arrive on Monday or Tuesday (01:13:25). Review of Meeting Closeout Summary: Joe Nayigiziki reviewed Claude's generated meeting closeout summary on the terminal, which outlined participants, decisions such as focusing on one region at a time and building Cletus, and to-dos regarding Ector County records (01:14:17). Joe noted that some items required explicit confirmation regarding open decisions, but expressed appreciation for the application's ability to immediately organize meeting dialogue (01:15:15). App Feature Requests, Artifact Management, and Meeting Wrap-up: Joe Nayigiziki requested a new feature for the meeting application to stream the terminal view into the app to enable real-time interactive conversations similar to working directly with Claude (01:17:05) (01:19:49). Reid Valentine raised a concern about finding a home for their markdown files and test results rather than relying solely on email. Joe noted that the application stores artifacts, and Reid mentioned having an existing workflow for reading handwriting that they could build upon (01:18:08). Joe concluded the meeting, and Reid agreed to email test results once completed (01:19:49). You should review Gemini's notes to make sure they're accurate. Get tips and learn how Gemini takes notes How is the quality of these specific notes? Take a short survey to let us know your feedback, including how helpful the notes were for your needs.