<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>Projects on D. Sébastien Taggart</title>
    <link>https://www.sebastientaggart.com/categories/projects/</link>
    <description>Recent content in Projects on D. Sébastien Taggart</description>
    <generator>Hugo</generator>
    <language>en</language>
    <lastBuildDate>Tue, 03 Nov 2026 16:54:00 -0500</lastBuildDate>
    <atom:link href="https://www.sebastientaggart.com/categories/projects/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Commissure</title>
      <link>https://www.sebastientaggart.com/projects/commissure/</link>
      <pubDate>Tue, 03 Nov 2026 16:54:00 -0500</pubDate>
      <guid>https://www.sebastientaggart.com/projects/commissure/</guid>
      <description>A local Python substrate and recipes for agents working on client data and documents.</description>
      <content:encoded><![CDATA[<p><strong>Commissure</strong> helps a small team do document work that used to take a large team. Point it at thousands of PDFs, Word files, spreadsheets, or downloaded records. It reads what can be read, flags what can&rsquo;t, runs the same complex analysis across every item, and keeps a pointer from every result back to the page it came from. People still check the work, but Commissure makes checking possible.</p>
<p>It runs on one machine, keeps your data there, and is free to use under the MIT license.</p>
<p><!-- raw HTML omitted -->Named for the commissure, the tissue that connects the two halves of the brain.<!-- raw HTML omitted --></p>
<hr>
<h2 id="the-problem">The problem</h2>
<p>Offices are being asked to do the same review, analysis, and reporting with a fraction of the people. The documents didn&rsquo;t get fewer. The deadlines didn&rsquo;t move. And the analysts who left took with them the memory of everything the office had already read, decided, and said.</p>
<p>The usual options don&rsquo;t fit:</p>
<ul>
<li><strong>A person and a spreadsheet</strong> is accurate but slow, and the whole exercise is repeated the next time the documents change.</li>
<li><strong>An AI chat tool</strong> is fast but handles one document at a time, guesses when it can&rsquo;t read something, and can&rsquo;t show where an answer came from.</li>
<li><strong>A one-off script</strong> written by a developer (or a coding agent) works on last Tuesday&rsquo;s sample and quietly breaks on the real batch. Nobody can say afterward what it skipped.</li>
</ul>
<p>Commissure is the middle path: the speed of automation with the paper trail of a careful analyst.</p>
<hr>
<h2 id="two-examples">Two examples</h2>
<h3 id="thousands-of-public-comments-one-analyst">Thousands of public comments, one analyst</h3>
<p>When an agency proposes a rule, the public responds on regulations.gov. The responses are often long, technical, and written by people who know the subject well. Reading, categorizing, and summarizing them accurately used to take a team of subject matter experts. One federal policy office lost most of that team and kept the obligation.</p>
<p>A Commissure solution now downloads every comment on a docket, decodes the attachments, flags the ones that can&rsquo;t be read, categorizes and summarizes the rest, and produces the report. Local AI models do the reading, so no comment text leaves the machine. Before the report goes out, the analyst reviews a random sample of the results against the original comments, and the measured error rate goes into the report alongside the findings.</p>
<h3 id="does-this-memo-conflict-with-anything-weve-already-said">Does this memo conflict with anything we&rsquo;ve already said?</h3>
<p>The same office checks new guidance for cross-cutting impacts. A memo about early childhood programs might touch mental health, Medicaid eligibility, or child care subsidies, and the office has to know whether it conflicts with, or is superseded by, guidance from three other agencies. That knowledge used to live in the heads of people who had read every memo for a decade.</p>
<p>The office loaded its archive of memos (a good share of them scans) into Commissure. When a new document arrives, Commissure finds the existing memos that overlap it and shows the passages side by side. An analyst makes the call on whether it&rsquo;s a conflict. The hours that went into finding are gone; the hours that go into deciding are the ones that were always worth spending.</p>
<hr>
<h2 id="what-commissure-does">What Commissure does</h2>
<ul>
<li><strong>Checks readability first.</strong> Every PDF, Word file, and spreadsheet is inspected before anything else happens. Scanned, corrupt, or image-only pages are set aside with a stated reason, or sent through local OCR, never silently skipped.</li>
<li><strong>Runs the same analysis on every document.</strong> The questions, checks, and extractions are defined once and applied consistently across the whole set.</li>
<li><strong>Keeps a pointer from every result to its source.</strong> Each extracted fact records the page and passage it came from, in the exact version of the file that was ingested.</li>
<li><strong>Makes the archive answerable.</strong> Load a body of documents once and ask it questions as new material arrives: what overlaps this, what supersedes this, what have we already said.</li>
<li><strong>Lets people verify the work.</strong> A reviewer grades a random sample of results against the originals. The resulting error rate is published with the deliverable, so readers know how far to trust it.</li>
<li><strong>Reruns only what changed.</strong> When documents, parameters, or code change, only the affected work is redone, and the differences are reported.</li>
<li><strong>Ships as a finished solution.</strong> A completed solution is packaged so it can be handed to another team or client and run as-is.</li>
</ul>
<hr>
<h2 id="why-commissure">Why Commissure?</h2>
<ul>
<li><strong>It lets fewer people do the work.</strong> Not by lowering the standard, but by moving the hours from finding and sorting to reviewing and deciding.</li>
<li><strong>Results you can defend.</strong> Every claim links to a source. The review process produces a real accuracy figure, not a vibe.</li>
<li><strong>Your data stays on your machine.</strong> No cloud service, no uploads by default. Sensitive material (PHI, PII, pre-decisional policy) goes to a remote model only if a solution explicitly allows it. Local models are supported for when it can&rsquo;t.</li>
<li><strong>AI where it helps, not everywhere.</strong> Predictable work is done by predictable code. AI handles only the judgment calls, and every one of those calls is logged.</li>
<li><strong>Built once, run again.</strong> A solution is a reusable recipe plus one client&rsquo;s data. Monthly reruns and handoffs are the normal case, not a rebuild.</li>
<li><strong>Cheap to own.</strong> MIT licensed, plain Python, standard SQLite. No per-seat fees, no vendor to call, nothing exotic for your team to maintain.</li>
<li><strong>Works with coding agents, under supervision.</strong> Agents follow the same rules the framework enforces, and their output goes through the same review as anyone else&rsquo;s.</li>
</ul>
<hr>
<h2 id="what-it-isnt">What it isn&rsquo;t</h2>
<ul>
<li><strong>Not a cloud service or a product you log into.</strong> It runs where you run it. Unattended runs use the operating system&rsquo;s own scheduler.</li>
<li><strong>Not an autonomous AI that works unsupervised.</strong> It produces evidence for people to judge. It does not rule on whether a memo conflicts with policy; it shows you the passages and lets you decide.</li>
<li><strong>Not a no-code tool.</strong> Building a new solution takes a developer, or a coding agent working with one. Running an existing solution does not.</li>
</ul>
<h2 id="links">Links</h2>
<ul>
<li><a href="https://github.com/sebastientaggart/commissure">Source on GitHub</a></li>
</ul>
]]></content:encoded>
    </item>
    <item>
      <title>Code Cannon</title>
      <link>https://www.sebastientaggart.com/projects/code-cannon/</link>
      <pubDate>Tue, 26 May 2026 00:30:51 -0400</pubDate>
      <guid>https://www.sebastientaggart.com/projects/code-cannon/</guid>
      <description>Skill syncing between AI agents — share capabilities across your agent fleet.</description>
      <content:encoded><![CDATA[<p>Code Cannon manages skills, prompts, and configurations across multiple AI coding agents. When you build a useful capability for one agent, Code Cannon lets you propagate it to your entire fleet — keeping your agents consistent and capable.</p>
<h2 id="what-it-does">What it does</h2>
<ul>
<li><strong>Skill synchronization</strong> — define skills once, deploy everywhere</li>
<li><strong>Agent configuration management</strong> — consistent settings across Claude, Cursor, and other tools</li>
<li><strong>Version-controlled capabilities</strong> — track what your agents can do and when it changed</li>
<li><strong>Convention-driven workflow</strong> — GitHub ticket integration and structured development patterns</li>
</ul>
<h2 id="links">Links</h2>
<ul>
<li><a href="https://github.com/lightbridgelab/code-cannon">Source on GitHub</a></li>
</ul>
]]></content:encoded>
    </item>
    <item>
      <title>OllamaMon</title>
      <link>https://www.sebastientaggart.com/projects/ollamamon/</link>
      <pubDate>Tue, 26 May 2026 00:30:51 -0400</pubDate>
      <guid>https://www.sebastientaggart.com/projects/ollamamon/</guid>
      <description>Monitor and benchmark your local Ollama instance — model performance at a glance.</description>
      <content:encoded><![CDATA[<p>OllamaMon gives you visibility into your local Ollama setup — which models are loaded, how much memory they&rsquo;re using, and how fast they&rsquo;re responding. Useful when you&rsquo;re running multiple models and want to know what&rsquo;s happening under the hood.</p>
<h2 id="what-it-does">What it does</h2>
<ul>
<li><strong>Live resource monitoring</strong> — VRAM/RAM usage per model</li>
<li><strong>Performance benchmarks</strong> — tokens/second for each loaded model</li>
<li><strong>Model management</strong> — see what&rsquo;s loaded, unload what you don&rsquo;t need</li>
<li><strong>Lightweight</strong> — runs as a simple CLI or background process</li>
</ul>
<h2 id="links">Links</h2>
<ul>
<li><a href="https://github.com/lightbridgelab/ollamamon">Source on GitHub</a></li>
</ul>
]]></content:encoded>
    </item>
    <item>
      <title>Deckhand</title>
      <link>https://www.sebastientaggart.com/projects/deckhand/</link>
      <pubDate>Thu, 30 Apr 2026 11:28:18 -0400</pubDate>
      <guid>https://www.sebastientaggart.com/projects/deckhand/</guid>
      <description>Open source bridge connecting AI agents to a physical Stream Deck for tactile, real-time control.</description>
      <content:encoded><![CDATA[<p>Deckhand bridges the gap between AI agents and physical hardware — giving you tactile, one-button access to agent actions through a Stream Deck or compatible device. Press a button, trigger an agent workflow.</p>
<h2 id="what-it-does">What it does</h2>
<ul>
<li><strong>Physical controls for digital agents</strong> — map Stream Deck buttons to agent commands</li>
<li><strong>Real-time status display</strong> — see agent state on your device&rsquo;s LCD keys</li>
<li><strong>Open source and extensible</strong> — works with any Stream Deck-compatible hardware</li>
<li><strong>Low latency</strong> — direct connection, no cloud round-trips</li>
</ul>
<h2 id="links">Links</h2>
<ul>
<li><a href="https://github.com/lightbridgelab/deckhand">Source on GitHub</a></li>
</ul>
]]></content:encoded>
    </item>
    <item>
      <title>NameForge</title>
      <link>https://www.sebastientaggart.com/projects/nameforge/</link>
      <pubDate>Tue, 26 May 2026 00:30:52 -0400</pubDate>
      <guid>https://www.sebastientaggart.com/projects/nameforge/</guid>
      <description>Find the perfect project name and matching available domain in seconds.</description>
      <content:encoded><![CDATA[<p>NameForge solves the annoying problem of naming a project — it generates creative, memorable names and immediately checks domain availability so you don&rsquo;t fall in love with something you can&rsquo;t register.</p>
<h2 id="what-it-does">What it does</h2>
<ul>
<li><strong>AI-powered name generation</strong> — describe your project, get dozens of brandable suggestions</li>
<li><strong>Instant domain checking</strong> — see .com, .dev, .io availability in real-time</li>
<li><strong>Favorites and export</strong> — save the ones you like, export for later</li>
<li><strong>No account required</strong> — just use it</li>
</ul>
<h2 id="links">Links</h2>
<ul>
<li><a href="https://github.com/lightbridgelab/nameforge">Source on GitHub</a></li>
</ul>
]]></content:encoded>
    </item>
    <item>
      <title>BookGrabberBot</title>
      <link>https://www.sebastientaggart.com/projects/bookgrabberbot/</link>
      <pubDate>Tue, 26 May 2026 00:30:53 -0400</pubDate>
      <guid>https://www.sebastientaggart.com/projects/bookgrabberbot/</guid>
      <description>Find, download, and manage e-books for the whole family from a single interface.</description>
      <content:encoded><![CDATA[<p>BookGrabberBot automates the tedious process of finding e-books, organizing them with proper metadata, and making them available to your family&rsquo;s reading devices. One search, and it handles the rest.</p>
<h2 id="what-it-does">What it does</h2>
<ul>
<li><strong>Automated e-book discovery</strong> — search multiple sources from one place</li>
<li><strong>Metadata tagging</strong> — proper titles, authors, covers applied automatically</li>
<li><strong>Family library management</strong> — organize books by reader, sync to devices</li>
<li><strong>Self-hosted</strong> — runs on your own hardware, no subscriptions needed</li>
</ul>
<h2 id="links">Links</h2>
<ul>
<li><a href="https://github.com/lightbridgelab/bookgrabberbot">Source on GitHub</a></li>
</ul>
]]></content:encoded>
    </item>
    <item>
      <title>MonitorPI</title>
      <link>https://www.sebastientaggart.com/projects/monitorpi/</link>
      <pubDate>Tue, 26 May 2026 00:30:53 -0400</pubDate>
      <guid>https://www.sebastientaggart.com/projects/monitorpi/</guid>
      <description>A Raspberry Pi setup with motion-sensing display for ambient dashboards and household information.</description>
      <content:encoded><![CDATA[<p>MonitorPI turns a Raspberry Pi with a touch screen into a smart ambient display — it wakes up when you walk by and shows useful information (weather, calendar, system status, whatever you configure), then goes back to sleep.</p>
<h2 id="what-it-does">What it does</h2>
<ul>
<li><strong>Motion-activated display</strong> — screen turns on when it detects presence, off when you leave</li>
<li><strong>Configurable dashboards</strong> — show whatever&rsquo;s useful for your household</li>
<li><strong>Simple setup script</strong> — one command to configure a fresh Pi</li>
<li><strong>Low power</strong> — designed to run 24/7 without noticeable electricity cost</li>
</ul>
<h2 id="links">Links</h2>
<ul>
<li><a href="https://github.com/lightbridgelab/monitorpi">Source on GitHub</a></li>
</ul>
]]></content:encoded>
    </item>
    <item>
      <title>Prompt Studio</title>
      <link>https://www.sebastientaggart.com/projects/prompt-studio/</link>
      <pubDate>Tue, 26 May 2026 00:30:54 -0400</pubDate>
      <guid>https://www.sebastientaggart.com/projects/prompt-studio/</guid>
      <description>A Mac-native app for managing, versioning, and evaluating AI prompts. Local-first, privacy-respecting.</description>
      <content:encoded><![CDATA[<p>Prompt Studio is a dedicated workspace for prompt engineering — designed for people who treat prompts as first-class artifacts rather than throwaway text. It runs entirely on your Mac with no cloud dependency, keeping your prompts private by default.</p>
<h2 id="what-it-does">What it does</h2>
<ul>
<li><strong>Versioned prompt library</strong> — organize, tag, and track changes to prompts over time</li>
<li><strong>Tone and output profiles</strong> — test the same prompt with different personas and output formats</li>
<li><strong>Evaluation suites</strong> — run structured tests to compare prompt performance</li>
<li><strong>Completely offline</strong> — no accounts, no telemetry unless you opt in</li>
<li><strong>Native macOS experience</strong> — built in Swift, feels like it belongs on your Mac</li>
</ul>
<h2 id="links">Links</h2>
<ul>
<li><a href="https://github.com/lightbridgelab/prompt-studio">Source on GitHub</a></li>
</ul>
]]></content:encoded>
    </item>
  </channel>
</rss>
