medpodgp / v2 /USER_GUIDE.md
nsugar's picture
MedPod GP 3.0.0
7d4d259 verified
|
Raw History Blame Contribute Delete
36.1 kB

MedPod GP - User Guide

Version 1.2 | CloudKites AI Lab


Table of Contents

  1. Introduction
  2. Who Is This For
  3. Getting Started
  4. Logging In
  5. Navigating the Application
  6. MedAudio - Speech-to-Text and Clinical Actions
  7. MedText - Clinical Text Assistant
  8. MedVision - Medical Image Analysis
  9. MedSearch - Curated Knowledge Search
  10. Medical Note Drafting
  11. Settings
  12. Data Storage and Privacy
  13. System Tray and Window Management
  14. Minimum System Requirements
  15. Troubleshooting
  16. Support and Contact

1. Introduction

MedPod GP is a desktop clinical AI assistant designed for general practice clinicians. It provides on-device speech-to-text transcription, clinical text analysis, and structured medical note generation - all running entirely on your local computer with no internet connection required.

Key Highlights

  • Completely offline - All AI processing happens on your computer. No patient data ever leaves your machine.
  • GPU accelerated - Optional NVIDIA CUDA support for ~10-15x faster inference. Works without a GPU too (CPU-only mode).
  • Speech-to-text - Record patient consultations and get instant transcriptions.
  • Clinical AI analysis - Generate summaries, differential diagnoses, clinical advice, and follow-up questions from transcripts.
  • Structured note generation - Produce EMR-ready clinical notes from your draft content.
  • Professional dark mode interface - Designed for clinical environments with reduced eye strain.

AI Models

MedPod GP ships with purpose-built AI models:

Model Purpose Size
MedGemma 1.5 4B Clinical text assistant (MedText) and medical image analysis (MedVision) - a medical-domain multimodal model tuned for general practice 4 billion parameters
Qwen3-ASR 0.6B Automatic speech recognition (MedAudio) - converts spoken audio to text 0.6 billion parameters
Qwen3.5 0.8B Auto-summariser - condenses long content before LLM processing 0.8 billion parameters

Models are loaded once at startup and remain in memory for fast, repeated use throughout your session.


2. Who Is This For

MedPod GP is built for healthcare professionals working in general practice and primary care settings, including:

  • General Practitioners (GPs) - for recording, transcribing, and documenting patient consultations
  • Practice Nurses - for capturing clinical notes during patient interactions
  • Medical Registrars and Trainees - for learning support with differential diagnosis suggestions
  • Allied Health Professionals - for documentation assistance in clinical settings

The application is intended as a clinical decision support tool. All AI-generated content should be reviewed and verified by a qualified healthcare professional before being used in patient care or medical records.


3. Getting Started

Application Installation

  1. Run the MedPod GP Installer (MedpodInstaller.exe) provided by your administrator.
  2. Follow the on-screen prompts. The default installation directory is:
    • Windows: C:\Program Files\MedPodGP
  3. The installer creates a desktop shortcut, a Start Menu entry under MedPodGP, and an uninstall shortcut.
  4. Tick "Launch MedPod GP" on the final page to start the app immediately.

AI Model Installation

MedPod GP requires AI model weights to be installed separately. Three installers are provided — one per model — all signed by CLOUDKITES PTY LTD:

Installer Model Required For Size
MedpodModels-medgemma.exe MedGemma 1.5 4B + vision projector MedText, MedVision, Note Drafting ~3.0 GB
MedpodModels-asr.exe Qwen3 ASR 0.6B MedAudio (speech-to-text) ~1.4 GB
MedpodModels-summariser.exe Qwen3.5 0.8B summariser Auto-summarisation of long content before LLM prefill ~540 MB

Double-click each installer in turn and click Next → Install. Each one automatically extracts its model into %LOCALAPPDATA%\medpod\models\ — no path picker, no component choices. All three are required for full functionality; at minimum install MedpodModels-medgemma.exe and MedpodModels-asr.exe.

Custom Model Location

To use models from a different directory, set the Windows environment variable MEDPODGP_MODEL_PATH to the full path of your models folder. The folder must contain the model subdirectories (medgemma1.5, qwen3-asr-0.6b, qwen3.5-0.8B).

First Launch

  1. Launch MedPod GP from your Start Menu or desktop shortcut.
  2. The application window appears immediately while AI models load in the background.
  3. You will see the login screen first.
  4. The ASR model (for speech-to-text) loads first, then the summariser, then MedGemma (for text analysis and vision).
  5. A status indicator shows the loading progress on each tab.

4. Logging In

When MedPod GP starts, you are presented with the login screen.

Sign In

  1. Enter your username in the first field.
  2. Enter your password in the second field.
  3. Click Sign In or press Enter on your keyboard.
  4. Your credentials are verified against the local SQLite database with salted SHA-256 password hashing.
  5. If authentication succeeds, you are taken to the main application.

Default Owner Account

On first run, MedPod GP creates a default owner account. Your system administrator will provide the initial credentials. The owner can create additional user and admin accounts from the Admin Dashboard.

User Roles

Role Permissions
Owner Full access. Can manage all accounts (users and admins), view system logs, and access all features.
Admin Can manage user accounts, view system logs, and access all clinical features.
User Access to all clinical features (MedAudio, MedText, MedVision, Notes). Cannot manage other accounts.

Exit from Login

Click the Exit button on the login screen to close the application entirely.

Logging Out

Once inside the application, click Logout in the top-right user profile section of the navigation bar. This returns you to the login screen. Any unsaved notes are automatically saved to the database. Your AI models remain loaded in memory, so logging back in is instant.

Auto-Logout

If an auto-logout timer is configured in Settings (1-480 minutes), the application automatically logs you out after the specified period of inactivity. Unsaved notes are auto-saved to the database before logout.


5. Navigating the Application

Navigation Bar

The top of the application contains a navigation bar with:

  • App logo ("MedPod GP") on the left
  • Tab buttons in the centre: MedAudio, MedText, MedVision, MedSearch
  • User profile section on the right: your avatar (first letter of your username), username, Logout button, and Settings (gear icon)

Click any tab button to switch between sections. The active tab is highlighted in green.

Title Bar

Below the navigation bar is a title bar showing the current section name with two toggle buttons on the right side:

  • Note - toggles the note drafting panel
  • Saved Notes - toggles the saved notes panel (showing previously saved notes from the database)

Layout

The main content area uses a split-panel layout:

  • Left panel - The active tab content (MedAudio, MedText, etc.)
  • Right panel - The Note panel or Saved Notes panel (hidden by default)
  • When you toggle either panel, the layout automatically adjusts to a 50/50 split so all columns are visible
  • You can drag the divider between panels to resize them.

Currently Available Tabs

Tab Status Description
MedAudio Active Audio recording, transcription, and clinical actions
MedText Active Clinical text assistant for questions and analysis
MedVision Active Medical image analysis with AI-generated reports
MedSearch Active Curated knowledge search (eTG, AMH, foundation guidelines) with Ask LLM

MedAudio is the default tab when you log in.


6. MedAudio - Speech-to-Text and Clinical Actions

MedAudio allows you to record patient consultations, transcribe them to text, and perform clinical analysis on the transcript.

Recording Audio

  1. Click the Record button (red) to start recording from your computer's microphone (or the device selected in Settings > Audio Input).
  2. While recording:
    • A recording duration timer displays the elapsed time.
    • A real-time spectrogram (48-bar visualiser) shows audio levels with a colour gradient from green (quiet) to yellow (moderate) to red (loud).
    • An audio level meter (progress bar) shows the current recording volume.
    • The status label shows "Recording...".
  3. Click the Stop button to end the recording.

Transcribing Audio

  1. After stopping a recording, click the Transcribe button.
  2. An animated processing overlay appears while the AI processes the audio. The interface remains responsive during this time.
  3. The transcribed text appears in the transcript area, streaming word by word.
  4. A green progress bar sweeps back and forth while tokens are being generated.
  5. When complete, the status shows "Transcription complete" with the processing speed.

Tip: For recordings longer than 10 minutes, increase the ASR Max Tokens value in Settings > Model Parameters. The default of 4096 supports approximately 30+ minutes of audio. The maximum setting of 8192 supports even longer recordings.

Clinical Action Buttons

After transcription, four action buttons appear below the transcript:

Button What It Does
Summarise Generates bullet-point medical notes suitable for a medical record from the transcript
Diagnosis Lists the top 5 differential diagnoses as a numbered list based on the transcript
Advise Provides clinical advice and recommended next steps based on the transcript
Further Questions Suggests follow-up questions to ask the patient to aid diagnosis and management

Clicking any action button:

  1. Automatically switches to the MedText tab.
  2. Sends the transcript along with the clinical instruction to MedGemma.
  3. Displays the AI-generated response in the MedText output area.

Cancelling

Click the Cancel button at any time to stop an ongoing transcription or clinical action. The interface resets and you can start a new recording.


7. MedText - Clinical Text Assistant

MedText is a chat-style interface for asking clinical questions and receiving AI-generated medical responses.

Asking a Question

  1. Type your clinical question in the text input field at the bottom of the MedText tab.
  2. Click the green Send button or press Enter to submit.
  3. An animated processing overlay appears while the model prepares the response.
  4. The response streams in token by token. A green progress bar indicates active generation.
  5. The final response is rendered with markdown formatting (headings, bullet points, bold text, etc.).

Example Questions

  • "What are the common causes of chest pain in a 45-year-old male?"
  • "List the first-line treatments for type 2 diabetes."
  • "What are the red flags for headache that require urgent referral?"
  • "Explain the CURB-65 scoring system for pneumonia."

Cancelling a Response

If the AI is generating a response and you wish to stop it:

  1. Click the red Cancel button (appears next to the Send button during generation).
  2. The generation stops immediately and the context is reset.
  3. You can type a new question straight away.

Important Notes

  • Each question starts a fresh context. Previous conversation history is not carried forward.
  • MedGemma is tuned for general practice clinical assistance. It will politely decline questions outside the healthcare domain.
  • All responses should be verified by a qualified healthcare professional.

8. MedVision - Medical Image Analysis

MedVision allows you to upload clinical images and receive AI-generated structured reports. It uses MedGemma 1.5 4B's vision pipeline (SiglipVisionEncoder + MultiModalProjector) to analyse the image alongside a text prompt.

Supported Imaging Types

Imaging Type Description
X-ray / CT / MRI Radiology images with body location selection (Head and Neck, Chest, Abdomen, Pelvis, Hand, Arm, Foot, Leg)
Mammography Breast imaging with BI-RADS assessment
Fundoscopy Retinal and optic disc evaluation
Skin Dermatological assessment with ABCDE criteria
Pathology Histopathology analysis
Endoscopy Gastrointestinal endoscopy findings
Ultrasound Ultrasound image interpretation
ECG Electrocardiogram interpretation
EEG Electroencephalogram interpretation
Custom Free-text query about any medical image

Uploading an Image

  1. Click the Upload Image button in the MedVision tab.
  2. A file dialog opens — select a JPEG or PNG image file.
  3. The image is displayed immediately in the image preview area. Click the image to open it in a full-screen image viewer with zoom (scroll wheel, +/- keys), pan (click-and-drag), and fit-to-window (F key). Press Esc to close the viewer.
  4. A thin animated progress bar appears while the image is encoded through the vision model (~5 seconds).
  5. When encoding completes, the status shows "Image encoded — ready for analysis".

Analysing an Image

  1. Select the Imaging Type from the dropdown (e.g., "X-ray / CT / MRI").
  2. If "X-ray / CT / MRI" is selected, an additional Body Location dropdown appears — choose the relevant region.
  3. Click the Analyse Image button.
  4. An animated processing overlay appears while the model generates the response.
  5. The AI-generated report streams token by token in the response area on the right.
  6. The final report is rendered with markdown formatting.

Custom Queries

For the "Custom" imaging type, or to ask a follow-up question about the uploaded image:

  1. Type your question in the text input field below the response area.
  2. Click the green Send button or press Enter.
  3. The AI analyses the image in the context of your specific question.

Cancelling

Click the Cancel button at any time to stop an ongoing analysis.

Important Notes

  • Each analysis starts with a fresh context. Previous analyses are not carried forward.
  • The vision encoder loads alongside MedGemma at startup. The status bar indicates when the vision pipeline is ready.
  • Image encoding happens in the background — the interface remains responsive during encoding.
  • MedVision uses MedGemma 1.5 4B as its underlying model. The same clinical AI capabilities apply to image analysis.

Adding Vision Results to Notes

Right-click on the MedVision response area to add findings to the Notes panel. The context menu options change based on the selected imaging type. For example, radiology reports offer sections like Findings, Technique, Impression, and Recommendations, while dermatology offers Description, Morphology, and Differential Diagnosis.


9. MedSearch - Curated Knowledge Search

MedSearch is a built-in clinical reference search across three curated knowledge packs that ship with the installer. Everything runs offline — no internet connection is needed, no patient query ever leaves your computer.

Knowledge Packs

Three packs are attached automatically at startup:

Pack Contents Size
eTG Therapeutic Guidelines (Australian) — prescribing guidance across antibiotic, respiratory, cardiovascular, psychotropic, paediatric, oncology, and toxicology topics. ~49 MB
AMH Australian Medicines Handbook — drug monographs with Mechanism, Indications, Precautions, Adverse effects, Dosage, and Products (brand-name listings). ~17 MB
foundation Additional guidelines — paediatric CPGs (RCH), Patient Blood Management, National Cervical Screening Program, COSA AYA, RANZCOG, and more. ~19 MB

The status line under the search bar shows "N pack(s) attached · M chunks indexed" once startup completes.

Searching

  1. Type a query into the search bar. The pack is searched with full-text indexing — exact phrases ("first-line treatment"), multiple keywords (paediatric anaphylaxis), or a single drug name (mebendazole) all work.
  2. Suggestions appear under the bar as you type, drawn from the autocomplete index of each pack.
  3. Click Search (or press Enter) to run the query.
  4. Results appear as a single-column list of cards, ranked by BM25F relevance. Each card shows:
    • The title (clickable — opens the full document)
    • The heading path in italics (e.g. "AMH > Mebendazole > Dosage")
    • A short body snippet with the matching context

Opening a Result

Click a result's title to open the full curated document. The doc viewer renders the content like a typeset journal page:

  • White page with serif body text and a centred reading column (~760 px wide)
  • Tables shown with a brand-green header bar (white labels), hairline grey grid, and a bold black bottom rule
  • A single citation footer at the bottom links to the original publisher's web page for the source

The content shown is MedPod's rewritten version of each chunk — the original publisher's text is never displayed inside the app. A small "Source:" line at the bottom shows the original reference for attribution.

Ask LLM (Grounded Q&A)

Instead of reading a single document, you can ask a question and let the LLM synthesise an answer from the top hits:

  1. Type a question in the search bar.
  2. Click the green Ask LLM button.
  3. MedSearch picks the top six hits, builds a context block, and streams a grounded answer.
  4. Each fact in the answer carries a numbered citation chip like [S1]. Click a chip to jump to the cited source.

Behind the scenes, every chip is verified against the cited snippet (exact / normalised / fuzzy match). Chips that can't be verified are visibly demoted so you know which claims are well-grounded and which to double-check.

Going Back

A ← Back button next to the search bar returns from the doc viewer or Ask LLM answer to the results list.

When Search Returns Nothing

If MedSearch reports "0 packs attached", the slim .sqlite packs aren't where MedPod expects them. By default they're installed at %LOCALAPPDATA%\medpod\packs\ by the MedPod installer. If you ran a custom install or deleted them, copy foundation.sqlite, etg.sqlite, and amh.sqlite back into that folder and restart the app.


10. Medical Note Drafting

The Notes panel allows you to build structured clinical notes from content across MedAudio, MedText, and MedVision, then generate a polished EMR-ready final note. Notes can be saved to the local database for later retrieval.

Opening the Notes Panel

Click the Note button in the title bar (right side). The panel appears to the right of the main content, automatically taking 50% of the window width. Click again to hide it. The button turns green when the panel is open.

Adding Content to the Draft

You can add content from any text area (MedText responses, MedAudio transcripts, or MedVision reports):

  1. Right-click on the text area containing the content you want to capture.
  2. A context menu appears with section options appropriate to the current tab:
    • MedAudio / MedText: Transcript, Summary, Differential Diagnoses, Investigation, Management Plan, Examination, Medical History, Advice, Further Questions
    • MedVision: Context-sensitive options based on the selected imaging type (e.g., Findings, Technique, Impression, Recommendations for radiology; Description, Morphology, Differential Diagnosis for dermatology)
  3. Select the appropriate section.

Text selection behaviour:

  • If you have highlighted text before right-clicking, only the selected text is captured.
  • If no text is highlighted, the entire content of the text area is captured with formatting preserved.

The Notes panel opens automatically (if hidden) and shows "Added to note draft".

If you add content to the same section more than once, the new content is appended to the existing entry.

Draft View

The Draft tab shows all captured sections formatted with markdown headings and separators. The draft is editable - you can manually refine the text before generating a final note.

The placeholder text reads: "Right-click on any text area to add content..."

Generating a Final Note

  1. Switch to the Notes panel and ensure you have draft content.
  2. Click the Generate Final Note button (green, at the bottom of the panel).
  3. The view switches to the Final Note tab automatically.
  4. An animated processing overlay appears while the active LLM generates the note.
  5. The structured clinical note streams in with standard EMR headings:
    • Presenting Complaint
    • History of Presenting Illness
    • Examination Findings
    • Investigation
    • Differential Diagnosis
    • Management Plan
    • Follow-up

The final note is editable - you can manually adjust any section before exporting.

Note: If the LLM is already generating content elsewhere (e.g. a MedText query), a "Model Busy" dialog appears offering to stop the current generation or continue waiting.

Switching Between Draft and Final Note

Use the Draft and Final Note tab buttons at the top of the Notes panel to switch between views.

Saving Notes to Database

  1. Click the Save button in the Notes panel.
  2. A dialog appears with fields for Title and Patient ID.
  3. Fill in the details and click OK to save to the local database.
  4. If the note was previously loaded from the database, the existing record is updated in-place (no duplicate is created).

Saved Notes Panel

Click the Saved Notes button in the title bar to view all previously saved notes.

The panel displays a table with columns:

Column Description
Title The note title entered when saving
Patient The patient ID associated with the note
Created Date and time the note was first saved
Updated Date and time of the last update
Actions Load and Del buttons for each row
  • Load - loads the saved note (both draft and final content) into the Notes panel for viewing or further editing. If you have unsaved changes in the current note, you are prompted to save first.
  • Del - permanently deletes the saved note from the database after confirmation.

Exporting Notes to File

Action Button Description
Copy to Clipboard Copy Copies the currently visible note (draft or final) to your clipboard for pasting into your EMR system
Save to Database Save Saves the note to the local SQLite database with title and patient ID
Export to File Export Opens a file dialog to save as Markdown (.md), plain text (.txt), Word (.docx), or PDF (.pdf)
Clear All Clear Removes all captured draft sections and resets the panel (with unsaved changes protection)

When exporting to Word or PDF, you can optionally include a doctor details header (configured in Settings > Profile tab) containing your name, qualifications, registration number, practice name, and contact information.

Unsaved Changes Protection

The application tracks whether you have modified the draft or final note content. If you attempt to:

  • Clear the notes - a dialog asks "Save & Clear", "Discard", or "Cancel"
  • Exit the application - a dialog asks "Save & Exit", "Exit Without Saving", or "Cancel"
  • Load a different saved note - you are prompted to save the current note first

Typical Workflow — Consultation

  1. Record a patient consultation in MedAudio.
  2. Click Transcribe to get the text.
  3. Click Summarise to generate clinical notes in MedText.
  4. Right-click the transcript in MedAudio and select Add to Note: Transcript.
  5. Right-click the summary in MedText and select Add to Note: Summary.
  6. Use Diagnosis, Advise, and Further Questions buttons as needed, adding each result to the appropriate note section.
  7. Review and edit the draft in the Notes panel.
  8. Click Generate Final Note to create a structured EMR-ready note.
  9. Review and edit the final note.
  10. Click Save to store in the database, Copy to paste into your EMR system, or Export to save as a file.

Typical Workflow — Image Analysis

  1. Switch to the MedVision tab.
  2. Click Upload Image and select a clinical image (JPEG or PNG).
  3. Wait for the image encoding progress bar to complete.
  4. Select the appropriate Imaging Type from the dropdown.
  5. Click Analyse Image to generate the AI report.
  6. Right-click the report to add findings to the appropriate note sections.
  7. Switch to the Notes panel, review the draft, and generate the final note.
  8. Save the note to the database or export to a file.

11. Settings

Access settings by clicking the gear icon in the top-right user profile section of the navigation bar. Settings are organised into two tabs: System and Profile.

System Tab

Models

Setting Options Default
ASR Model (MedAudio) Qwen3-ASR 0.6B, Whisper Small, Whisper Medium Qwen3-ASR 0.6B
LLM Model (MedText) MedGemma 1.5 4B (Q4_K_M), Qwen3.5 4B (Q4_K_M), Qwen3.5 2B (Q4_K_M) MedGemma 1.5 4B
LLM Model (MedVision) MedGemma 1.5 4B (shared with MedText) MedGemma 1.5 4B
Search (MedSearch) Not yet configured -

Note: MedText LLM model changes take effect when you click Save — the model hot-swaps without restarting the application. MedVision always uses MedGemma 1.5 4B as it requires the vision encoder.

Audio Input

Setting Description Default
Microphone Select a specific audio input device, or use the system default System Default

Click Refresh to rescan for newly connected audio devices.

Model Parameters

Setting Range Default Description
LLM Max Tokens 128 - 4096 1024 Maximum length of AI-generated text responses (MedText, Final Note, clinical actions)
ASR Max Tokens 512 - 8192 4096 Maximum length of ASR transcription output; higher values support longer recordings (~30+ minutes at default)
Auto-Summarise Threshold Configurable Default threshold When input content exceeds this token count, it is automatically condensed by the Qwen3.5 0.8B summariser before being sent to the main LLM

Appearance

Setting Range Default
Content Font Size 10pt - 24pt 13pt
Theme Dark Mode (Light mode planned) Dark Mode

System Options

Setting Description Default
Minimize to system tray on close When enabled, closing the window hides it to the system tray instead of exiting Enabled
Desktop notifications for model events Show system notifications when models finish loading or other significant events occur Enabled
Auto-logout timer Automatically logs out after the specified number of minutes of inactivity (0 = disabled) 0 (disabled)

The auto-logout timer range is 0-480 minutes. When triggered, any unsaved notes are automatically saved to the database before returning to the login screen.

Profile Tab

The Profile tab allows you to configure your doctor details, which are used as an optional header when exporting notes to Word (.docx) or PDF (.pdf) format.

Field Description
Name Your full name and title (e.g., Dr. Jane Smith)
Qualifications Your medical qualifications (e.g., MBBS, FRACGP)
Registration Number Your medical registration number
Practice Name The name of your practice or clinic
Address Practice address
Phone Practice phone number
Email Practice email address

These details are saved per-user in the local database and persist across sessions.

Saving and Cancelling

  • Click Save to apply your changes and return to the application. Settings are saved to the local database.
  • Click Cancel to discard changes and return with previous settings intact.
  • Click the X button in the top-right corner of the settings card to close without saving.

12. Data Storage and Privacy

Local Database

All user data is stored in a local SQLite database on your computer:

  • Windows: %LOCALAPPDATA%\medpod\data\medpod.db

The database stores:

Data Description
Accounts Usernames, salted password hashes, roles, creation dates
Settings Per-user application settings (model preferences, appearance, system options)
Profiles Per-user doctor details for note export headers
Notes Clinical notes with title, patient ID, draft markdown, final markdown, and timestamps
Logs System event audit trail (logins, account changes)

Privacy

  • All data remains on your local computer. No data is transmitted to any server or cloud service.
  • The database file is not encrypted at the filesystem level — physical device security is managed by your facility's IT infrastructure.
  • Passwords are stored as salted SHA-256 hashes, never in plain text.
  • Deleting a user account removes all associated data (settings, profiles, notes, logs) from the database.

13. System Tray and Window Management

Minimize to Tray

When "Minimize to system tray on close" is enabled in Settings (the default):

  1. Closing the window (X button) hides it to the system tray instead of quitting.
  2. A notification appears: "Application minimized to system tray."
  3. The MedPod GP icon appears in your system tray (notification area).

Restoring the Window

  • Click or double-click the tray icon to bring the window back.
  • Right-click the tray icon and select Show from the context menu.

Exiting the Application

When minimize-to-tray is enabled:

  • Right-click the tray icon and select Exit to quit completely.

When minimize-to-tray is disabled:

  • Closing the window shows a confirmation dialog.
  • If you have unsaved notes, you are offered: Save & Exit (saves notes to database then exits), Exit Without Saving (discards unsaved changes), or Cancel (returns to the application).
  • If no unsaved notes, a standard exit confirmation is shown.

If AI inference is running when you exit, all running models are gracefully cancelled before the application closes.


14. Minimum System Requirements

Hardware

Component Minimum Recommended
Processor x86-64 CPU with AVX2 support AMD Ryzen 7 / Intel Core i7 (8th gen or later)
RAM 8 GB 16 GB or more
Storage 10 GB free space (for application and models) SSD recommended for faster model loading
GPU (optional) Not required for CPU builds NVIDIA GPU with 6+ GB VRAM (RTX 2060 or later) for CUDA builds
Microphone Any USB or built-in microphone -

Software

Component Requirement
Operating System Windows 10 (64-bit) or later; Linux (x86-64 with ALSA)
Display 1280 x 720 minimum resolution
GPU Driver (CUDA only) NVIDIA GPU driver (installed automatically with GeForce Experience / NVIDIA App)

Performance Reference

CPU only — tested on AMD Ryzen Z1 Extreme with 16 GB LPDDR5X 7500 MHz:

Task Performance
MedGemma 1.5 text generation ~5-6 tokens/second
MedVision image encoding ~5 seconds
MedVision report generation ~5-6 tokens/second
Qwen3-ASR speech-to-text ~14-16 tokens/second
Model loading (first launch) 30-60 seconds

CUDA GPU — tested on NVIDIA RTX 3090 (24 GB VRAM):

Task Performance
MedGemma 1.5 text generation ~68-97 tokens/second
MedVision image encoding ~0.5 seconds
MedVision report generation ~68 tokens/second
Qwen3-ASR speech-to-text ~600 tokens/second (with CUDA Graph)
Model loading (first launch) 30-60 seconds

The CPU build requires no GPU — all inference runs on the CPU. The CUDA build requires only an NVIDIA GPU with drivers installed (no CUDA Toolkit or developer tools needed — all runtime libraries are bundled with the application).


15. Troubleshooting

Application does not start

  • Ensure your system meets the minimum requirements.
  • Check that the installation completed successfully.
  • Verify that the model files are present in the models directory (default: C:\Program Files\MedPodGP\models). Each model subdirectory (medgemma1.5, qwen3-asr-0.6b, qwen3.5-0.8B) must contain its weight files and tokenizer.

Models take a long time to load

  • Model loading happens once at startup. The ASR model loads first, then MedGemma.
  • On systems with slower storage (HDD), loading may take longer. An SSD is recommended.
  • The application is fully usable while models are still loading - the relevant tab will show a loading indicator.

Audio recording does not work

  • Ensure your microphone is connected and set as the default recording device in your OS sound settings.
  • Try selecting a specific microphone in Settings > Audio Input instead of "System Default". Click Refresh to rescan for devices.
  • Check that no other application is exclusively using the microphone.
  • Try unplugging and reconnecting your microphone.
  • On Linux, ensure ALSA is available (libasound.so.2). PipeWire with ALSA compatibility also works.

Transcript is truncated for long recordings

  • Increase the ASR Max Tokens value in Settings > Model Parameters. The default of 4096 supports ~30+ minutes. Set it higher (up to 8192) for longer recordings.

AI responses are slow

  • AI inference speed depends on your CPU. More powerful processors generate tokens faster.
  • Close other CPU-intensive applications to improve performance.
  • Consider reducing the LLM Max Tokens setting in Settings to get shorter, faster responses.

The application is not responding during generation

  • The interface should remain responsive during all AI operations. If you see a spinning overlay or a green progress bar, the model is working normally.
  • You can always click Cancel to stop an ongoing generation.

Exit while model is running

  • The application gracefully cancels all running AI operations when you close it. No action is needed.

16. Support and Contact

CloudKites AI Lab

Address Sydney, NSW, Australia
Website https://cloudkites.com
Email contact@cloudkites.com

For technical support, bug reports, or feature requests, please contact us via email at contact@cloudkites.com.


MedPod GP v1.2 - Built by CloudKites AI Lab, Sydney, Australia. All AI inference runs locally on your device. No patient data is transmitted over the internet.