A local-first clinical knowledge system

Every guideline. One private AI bot.

ESC and ACC/AHA recommendations, linked into a living knowledge graph and queryable from my phone—with no subscriptions or usage-based AI token charges.

Runs locallyCites every answerVoice enabledPrivate by design
2 societiescompared side by side
1 noteper recommendation
<10 secvoice to cited answer
1 weekendfrom idea to working bot
01 · The real problem

The bottleneck was never understanding guidelines. It was retrieving the right recommendation at the moment I needed it.

📚

Thousands of graded recommendations

Different societies, formats, update cycles, classes and levels of evidence.

⚖️

Meaningful disagreement

DAPT duration, lipid targets, revascularisation thresholds and invasive strategy timing.

⌛

PDF friction at the bedside

Open, search, scroll, cross-reference—then repeat in another guideline.

02 · The knowledge engine

Guidelines in. Cited answers out.

Four deliberate transformations turn static PDFs into an assistant that can retrieve, compare and explain—without inventing clinical recommendations.

📄
Stage 01

Ingest

Convert source PDFs to Markdown while preserving figures, tables and central illustrations.

✂️
Stage 02

Atomize

Split each guideline into one note per recommendation; lift class and evidence into metadata.

🕸️
Stage 03

Connect

Link parallel recommendations, clinical concepts, disagreements and personal case notes.

💬
Stage 04

Query

Ask by text or voice. The bot retrieves local notes and returns a concise, cited response.

VOICE QUERY · 08:41
“For an older patient with NSTE-ACS and high bleeding risk, how do ESC and ACC/AHA positions differ on antiplatelet duration?”
◎ Answer grounded in 4 linked notes · sources shown
Deterministic ingestionPattern matching extracts recommendations; the model does not rewrite them.
Retrieval before generationThe bot answers from the vault or says the evidence is absent.
Human verificationEvery statement links back to the source note and original guideline.
03 · Build sequence

Prove the loop, then scale it.

The safest order is the one that reveals weak links early. Open each stage for the practical logic.

Install the local model runner and agent, connect Telegram, restrict access to your user ID, and prove a round trip before adding content.
ollama pull qwen3:8b
ollama serve
hermes gateway setup
hermes gateway
Choose the guideline you know best. Convert it, split it, then compare twenty random notes against the original PDF. Record the error rate.
Recommendation tables vary. Tune extraction rules to the actual converted structure and always run a dry test before writing notes.
A single society is a reference. Two societies on the same topic create useful comparisons and reveal genuine divergence.
Lay out both positions, classes, evidence levels and pivotal trial evidence in your own words with direct citations.
Add guidelines steadily. Schedule one neglected recommendation to be pushed to your phone each day for deliberate review.
04 · The local-first stack

Six tools. One private system.

Each component has a single job, and your clinical source material remains on your own machine.

◆
Knowledge base

Obsidian

Plain Markdown notes, structured metadata, backlinks and graph view—stored locally.

The vault
✦
Agent layer

Hermes Agent

Reads the vault, orchestrates retrieval and connects the local model to messaging.

The operator
◉
Local inference

Ollama

Runs the language model on your own hardware. No hosted model account required.

The engine
⌁
Language model

Qwen 3

Open-weight models in several sizes so you can match speed and memory to your machine.

The reasoning layer
➤
Mobile interface

Telegram

A familiar text and voice front door, locked to one authorised user.

The interface
≋
Local speech

faster-whisper

Transcribes voice notes locally before retrieval begins. Audio stays on the machine.

The ears
05 · Trust boundaries

Useful because it is linked. Safe because it is bounded.

🔒

Private by architecture

The full guideline corpus never needs to be published or uploaded to a hosted model.

  • Local vault and local inference
  • One authorised messaging user
  • Local voice transcription
  • No usage-based AI token billing
©

Copyright by design

Freely downloadable clinical guidance is not automatically freely redistributable.

  • Full source text stays private
  • Published comparisons use original prose
  • Facts, classes and levels are cited
  • Commercial use needs qualified advice
Build your own specialty assistant

Turn static guidance into a living clinical memory.

The system retrieves evidence faster. It does not replace judgment, validate a diagnosis or make a clinical decision.

Start with Hermes ↗
Link copied