Blog
Thoughts on self-hosted infrastructure, local AI, automation, and building things that respect your privacy.
SolScribe 2.0 keeps what it knows current
SolScribe 2.0 changes what happens after a recording is transcribed. Brain checks each new fact against what it already holds.
Prism 1.10 answers from whole passages
Prism 1.10 reads the full passage behind each source when it answers a question, instead of a short excerpt.
DropSnip 1.3 makes annotations editable
DropSnip 1.3 lets you edit annotations after drawing them and takes Scrolling Capture out of beta. History now shows previews.
DeviceWatch now groups recalls by FDA recall event
DeviceWatch now shows one recall card per FDA recall event, with the FDA class where matching enforcement records have one.
SolScribe 1.9 brings a redesigned web app
SolScribe 1.9 ships a new web design. Home replaces the dashboard and Brain opens on a map of your projects.
DeviceWatch signals now link to the CAPAs in your eQMS
DeviceWatch can now link signals to the CAPAs that follow them up and point each CAPA at its record in your eQMS.
What FDA AEMS means for MAUDE searches and DeviceWatch
FDA announced that MAUDE data will move into AEMS. What that means for MAUDE searches today and how DeviceWatch is preparing.
SolScribe's Brain can now propose tasks for you to accept
SolScribe's Brain has a new opt-in task review mode. It proposes tasks from your recordings and nothing becomes a task until you accept it.
Self-hosted SolScribe now uses local AI models by default
Self-hosted SolScribe 1.8.7.0 uses local AI models unless you tell it otherwise.
SolScribe's Brain now keeps a status note for each project
SolScribe's Brain now tracks the projects you talk about and keeps a status note for each one, citing the recordings behind every line.
A one-page checklist for law firms vetting an AI tool
I published a one-page checklist for law firms deciding whether to put client files into an AI tool.
What a self-hosted SolScribe server sends, and to whom
SolScribe's server images no longer report usage to third parties. As of 1.8.6.0, the README lists every connection a self-hosted server makes.
Prism 1.8 gives better answers from the same sources
Prism 1.8 changes how search ranks results and how Ask builds its context. The sources are the same.
Alerts that start with the FDA's own coding
DeviceWatch now raises a critical alert as soon as a MAUDE report arrives coded by the FDA as a Death, without waiting for the narrative or AI analysis.
Better recording recovery in SolScribe
SolScribe now saves browser audio as you record and recovers interrupted uploads more reliably.
When the FDA publishes the narrative late
DeviceWatch now checks for MAUDE narratives the FDA publishes after the initial report and adds late text through the audit trail.
Parakeet learns medical vocabulary and one feature stays off
SolScribe's Parakeet engine now takes a custom vocabulary on each transcription. A speech detector for splitting long audio ships switched off.
Prism now finds apps and files by name
Prism 1.7.0 finds apps and files by name through Spotlight on macOS and Everything on Windows, and keeps those names out of AI.
FDA is putting Elsa on top of submission data. Now show the work.
FDA has confirmed less about its AI assistant's role in device review than the trade coverage suggests.
SolScribe's background work now survives a crash
Durable job queues replace fragile background tasks in SolScribe, with support for PostgreSQL and SQLite.
Prism's Ask now cites what it actually retrieved
How Prism 1.4 rebuilt retrieval, grounded citations and made connector failures visible.
MAUDE hasn't been replaced, and we said it had
Correcting DeviceWatch’s claims about MAUDE and the FDA’s planned AEMS transition.
Manual MAUDE surveillance doesn't scale
Read faster, hire another analyst or narrow the scope. Each delays the point where the method fails.
Why most competitor monitoring programs miss what matters
Competitor adverse events get collected every quarter and rarely change a decision. A framework that connects them to risk files, design reviews and PSURs.
MCP servers: teaching Claude to talk to my homelab
Three MCP servers let Claude search my Paperless documents, run n8n workflows and query my vector database. What I built and what I would repeat.
80 workflows later: how n8n replaced my SaaS stack
Most of the tools I paid for were workflow automation with a monthly invoice. What I rebuilt in n8n, the patterns that held up and what it saved.
SolScribe vs Cloud Transcription: Why Self-Hosted Wins
Cloud transcription services charge per minute, store your audio on their servers, and can change terms overnight. Here's how SolScribe delivers the same features on hardware you control.
How AI reads adverse event reports
What a language model can extract from a MAUDE narrative, how signal detection uses it and why a qualified reviewer still signs off.
The Digital Brain: a self-hosted personal knowledge management system
Notes, documents, bookmarks, voice memos and wiki pages flow into one searchable knowledge base. The four layers and the automation between them.
FDA MAUDE database: a working guide for medical device teams
Who reports to MAUDE, how to search it, where the data misleads and how to build surveillance around it.
My local LLM stack: private AI on an 8GB GPU
Language models, speech-to-text, embeddings and document search on a $300 graphics card. The setup, the cost and where it falls short.
Building a 24-service homelab on a $400 NAS
A TerraMaster F2-424 running TrueNAS and Docker replaced about $190 a month of subscriptions. What runs on it, what it costs and what I learned.
Preparing for FDA AEMS: a checklist for device teams
MAUDE data is due to move into AEMS and FDA has given no device date. Five things a surveillance team can do before it does.
Introducing DeviceWatch: AI-Powered FDA MAUDE Surveillance for Medical Device Teams
Automated adverse event monitoring, AI summarization, and compliance-ready reports — so your regulatory team can stop searching and start analyzing.