Quryltai AI

A privacy-first local AI meeting protocol assistant for HackAlem. The planned application turns Russian, Kazakh, and mixed-language meetings into speaker-labeled transcripts, action items, summaries, and DOCX/PDF protocols.

Final MVP status

The Streamlit application has three clearly separated paths:

• Demo is the original fictional, deterministic Russian/Kazakh example and works without Ollama.
• Text transcript analyzes pasted Russian, Kazakh, or mixed text using a locally running Ollama model. It produces grounded action items, source quotes, confidence scores, main topics, key problems, decisions, and an overall summary.
• Audio (Phase 3) transcribes WAV, MP3, or M4A locally with faster-whisper and sends the resulting text only to the same local Ollama pipeline.

DOCX and PDF downloads contain meeting metadata, participants, speaker mapping, action items, summary, and transcript. Uploaded audio is written only to an operating-system temporary file during transcription and is deleted immediately afterward. Speaker edits are reflected in displayed and exported results but are not persisted.

MVP architecture

src/llm/base.py defines a provider interface. src/llm/ollama.py implements it with the Ollama HTTP API, schema-constrained output, Pydantic validation, and loopback-only URL enforcement. src/understanding.py turns pasted lines into evidence-addressable segments, sends one structured request to the local model, and rejects unknown evidence IDs, non-verbatim quotes, invented assignees, and invented deadline phrases before building a Protocol.

The provider boundary leaves room for another self-hosted runtime later without changing meeting analysis or the UI. There is no cloud provider implementation and no cloud fallback.

Install (Windows PowerShell)

Use Python 3.11 or 3.12. The requirements include Streamlit, Pydantic, faster-whisper, python-docx, ReportLab, and pypdf. Installing them does not download an Ollama or Whisper model.

This checkout was verified with the already installed Python 3.13.15, Streamlit 1.64.0, and Pydantic 2.13.5. Its .venv is ready to run. A fresh installation can use the commands below.

py -3.12 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt

If Python is installed as python instead of py, use python -m venv .venv. With uv already installed, an alternative is:

uv venv --python 3.12 .venv
uv pip install --python .venv\Scripts\python.exe -r requirements.txt

Configure Ollama locally

Install Ollama from https://ollama.com/download/windows. The recommended laptop model is qwen3:4b-instruct-2507-q4_K_M, a roughly 2.5 GB Q4 model. It is small enough for a hackathon laptop and the Qwen3 family supports 100+ languages. This is a pragmatic starting point; evaluate Kazakh extraction on your own recordings. For a stronger machine, qwen3:8b is about 5.2 GB.

No model is downloaded by this repository. After installing Ollama, run these commands yourself:

[Environment]::SetEnvironmentVariable('OLLAMA_NO_CLOUD', '1', 'User')
$env:OLLAMA_NO_CLOUD = '1'
ollama pull qwen3:4b-instruct-2507-q4_K_M
ollama list

Restart the Ollama application after setting the persistent environment variable. OLLAMA_NO_CLOUD=1 disables Ollama cloud features. Quryltai additionally rejects non-loopback provider URLs.

Optional configuration must still point to loopback:

$env:OLLAMA_BASE_URL = 'http://127.0.0.1:11434'
$env:QURYLTAI_OLLAMA_MODEL = 'qwen3:4b-instruct-2507-q4_K_M'
$env:QURYLTAI_OLLAMA_TIMEOUT_SECONDS = '300'

Configure local audio transcription

Audio mode never downloads weights automatically. Point it at an existing local faster-whisper model folder:

$env:QURYLTAI_WHISPER_MODEL_PATH = 'C:\HackAlemModels\faster-whisper-small'

If you do not already have local Whisper weights and choose to download them, run this once while online (the small multilingual model is several hundred MB):

New-Item -ItemType Directory -Force C:\HackAlemModels | Out-Null
.\.venv\Scripts\hf.exe download Systran/faster-whisper-small --local-dir C:\HackAlemModels\faster-whisper-small

The current MVP transcribes speech but does not diarize multiple voices. Audio becomes neutral transcript segments; provide participant names as metadata and review the transcript before relying on assignments.

Run

$env:QURYLTAI_DATA_DIR = 'C:\HackAlemData'
$env:OLLAMA_BASE_URL = 'http://127.0.0.1:11434'
$env:QURYLTAY_OLLAMA_MODEL = 'qwen3:4b-instruct-2507-q4_K_M'
$env:QURYLTAI_OLLAMA_TIMEOUT_SECONDS = '300'
# Required only for Audio mode:
$env:QURYLTAI_WHISPER_MODEL_PATH = 'C:\HackAlemModels\faster-whisper-small'
.\.venv\Scripts\python.exe -m streamlit run app.py

Open http://127.0.0.1:8501. Leave Demo mode enabled for the offline deterministic demo. For real local understanding, turn Demo mode off, select Text transcript, paste the example below, and click Analyze meeting:

Айгерім: Тимур, подготовь бюджет до пятницы.
Тимур: Жақсы, жұмаға дейін дайындаймын.
Дана: Ещё нужно проверить договор, бірақ жауапты адам әлі жоқ.
Айгерім: Шешім қабылданды: пилотты бес дүкенмен бастаймыз.

Use one utterance per line with Name: text or [Name]: text. Lines without a prefix use the neutral speaker label Transcript. Relative dates are resolved against the meeting date and Asia/Qyzylorda timezone when unambiguous. For audio, select Audio (Phase 3), upload WAV/MP3/M4A, and analyze. After any successful mode, use the enabled Export DOCX and Export PDF download buttons.

The default fictional date is September 23, 2026. Participants are metadata, not automatic voice identification. Input changes apply when the form is submitted. The app preserves results through normal widget reruns; a new session/server restart loses them.

Presentation demo checklist

1. Start Ollama and verify ollama list shows qwen3:4b-instruct-2507-q4_K_M.
2. Start Streamlit with the commands above.
3. Keep Demo mode on, select Analyze meeting, edit one speaker name, and download both protocol formats. This path needs neither Ollama nor Whisper.
4. Turn Demo mode off, keep Text transcript, paste the short sample above, and analyze it with local Ollama.
5. Optional audio demo: configure QURYLTAI_WHISPER_MODEL_PATH, select Audio (Phase 3), upload a short clear WAV/MP3/M4A clip, and analyze it.

Privacy and local storage

• Demo processing runs inside Streamlit. Text mode connects only to the configured loopback Ollama service. No meeting-data path to OpenAI or another external/cloud AI API exists in the application.
• Set OLLAMA_NO_CLOUD=1, restart Ollama, and preload the model before entering a closed environment. Runtime analysis then needs no internet connection.
• Streamlit binds to 127.0.0.1, and usage statistics are disabled in .streamlit/config.toml. Run from the repository root so this configuration is loaded. Do not expose the development app through public tunnels or cloud hosting.
• Uploaded audio remains local. A temporary transcription copy is deleted immediately; meeting results remain in Streamlit session memory.
• The future data directory defaults to C:\HackAlemData on Windows (~/HackAlemData elsewhere). Override with QURYLTAI_DATA_DIR before launch. Phase 2 validates the location but does not create it or write meeting files.
• Repository paths, network shares, and known OneDrive locations are rejected. The user must verify the chosen directory is not synchronized by another backup/sync service. This repository itself may be in OneDrive; runtime meeting files must stay outside it.
• .gitignore excludes common audio, transcript, document, environment, model, and temporary file formats. It is a second guard, not a substitute for external storage; it does not protect already tracked files or forced Git additions.
• Package installation requires internet access unless using predownloaded packages. Runtime demo analysis can work offline. No database is used.

Structure

app.py                    Streamlit interface
src/config.py             External data path validation
src/schemas.py            Pydantic models and reference validation
src/demo.py               Fictional bilingual fixtures (safe source code)
src/orchestrator.py       Phase 1 deterministic demo workflow
src/understanding.py      Grounded text analysis and prompt
src/transcription.py      Local faster-whisper adapter
src/exporters.py          In-memory DOCX and PDF generation
src/llm/base.py           Local provider abstraction
src/llm/ollama.py         Loopback-only Ollama provider
.streamlit/config.toml    Loopback binding and telemetry opt-out
tests/test_foundation.py  Pipeline, schemas, privacy path checks
tests/test_app.py         Streamlit UI smoke checks
tests/test_understanding.py Structured-output and grounding checks
tests/test_exports.py     Audio adapter and real export checks
requirements.txt          Final MVP dependencies

Component status

1. normalize_audio: handled internally by faster-whisper/PyAV for supported uploads.
2. transcribe_audio: implemented with faster-whisper and explicitly configured local weights.
3. diarize_speakers: not implemented in the MVP; audio uses a neutral speaker label.
4. align_speakers: not implemented because diarization is not included.
5. extract_tasks: implemented for pasted text with local Ollama, validation, and evidence references.
6. generate_summary: implemented for pasted text with local Ollama.
7. generate_protocol: implemented as validated in-memory DOCX and PDF exports using Unicode-capable local fonts.

The deterministic demo remains available without Ollama or Whisper. There is no external cloud AI fallback.

Basic checks

.\.venv\Scripts\python.exe -m unittest discover -s tests -v

Checks use mocked providers and synthetic fixtures, require no model, make no network calls, and do not retain meeting files. Real transcription and extraction quality still depend on the selected local models and representative Russian/Kazakh recordings.