"Local intelligence for local clinics."
Offline clinical consultation, Taglish speech transcription & medication safety-review assistant.
AppBuildersPH Hackathon 2026 Submission · Theme: Local AI (On-Device Inference)
"An AI healthcare product that remains genuinely useful when the cloud disappears."
A local-first, air-gapped clinical documentation and medication safety workstation designed for Rural Health Units (RHUs), barangay health stations, and disaster-response medical missions across the Philippines. Runs entirely on an ordinary laptop or workstation; no cloud AI, no telemetry, no subscription bills. All patient health records are encrypted at rest with AES-256-GCM.
Under the Philippine Data Privacy Act of 2012 (DPA), transmitting patient consultations and identifiable medical records to third-party cloud servers creates severe compliance, ethical, and privacy risks. Furthermore, during typhoons, natural disasters, or in remote island/mountain RHUs with zero connectivity, cloud healthcare tools fail completely. Lokal.MD runs 100% on-device: patient data never leaves the hardware, operations never stop during blackouts, and public health units incur ₱0 in subscription fees.
- Speech-to-Text:
faster-whisper(Systran/faster-whisper-base/small) running locally via CTranslate2. - Clinical Structuring (SOAP Notes):
llama3.2:3brunning on-device via local Ollama. - Safety Engine: Deterministic offline SQLite rules engine checking contraindications, allergies, and dosage limits against the Philippine National Formulary (PNF) standards.
- AI Development Tools: Built using Devin (Cognition) + Antigravity.
Status: functional MVP. Ships with a synthetic sample medication reference set (labelled as such on every result and PDF). It is not a certified medical device and has not been assessed for Data Privacy Act or regulatory compliance. See "Limitations".
setup.bat # once, with internet: creates venv, downloads dependencies, Whisper & Ollama
run.bat # daily use, 100% offline; opens http://127.0.0.1:8501# 1. Setup virtual environment and dependencies
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
# 2. Prepare local models (one-time while online)
python scripts/prepare_models.py
# 3. Launch application (100% offline)
streamlit run app.py- First launch prompts you to create the clinic administrator account and securely presents a one-time master recovery key.
- Run automated tests:
pytest(94 passing tests). - Offline air-gap self-test:
python scripts/check_offline.py.
| Area | Implementation |
|---|---|
| Workspace | Intake → Transcript → SOAP note → Medication safety → Review & approve, one encounter view with unsaved-change tracking |
| Speech | faster-whisper from local files (models/whisper/<size>); WAV/MP3/M4A/OGG/FLAC upload or browser microphone; low-confidence lines marked [?] and must be cleared by a human; raw audio deleted after processing unless an admin enables retention with a policy reference |
| Notes | Ollama (default llama3.2:3b) constrained to a JSON schema; vitals, allergies, medicines and orders are copied by code, clinician entries override the AI, every AI item is grounding-checked (numbers + wording, Tagalog glossary) and unsupported items are withheld; malformed output → retry → template fallback; notes stay drafts until a clinician approves |
| Medication safety | SQLite-backed deterministic rules engine (rhuscribe/safety), six result categories, evidence + next step per finding, clinician acknowledgement required before approval |
| Storage | SQLite + per-field AES-256-GCM, scrypt-wrapped data key per user, recovery key, hash-chained audit log, encrypted backups, restore, retention, hard delete |
| Access | Roles: admin, clinician (only role that can approve), staff, pharmacist; purpose-of-access for history; inactivity lock (auto-saves drafts, wipes session) |
| Export | ReportLab PDF with DRAFT/APPROVED banners, medication findings and approval record exactly as stored (no invented signatures) |
- Offline by construction: Streamlit binds to 127.0.0.1; telemetry off; no external assets; a process-level socket guard
refuses any non-loopback connection; models load with
local_files_only; the Ollama client ignores proxies. - Encryption is not SQLite's: SQLite is unencrypted; clinical columns are encrypted by the app. Encounter IDs, timestamps, status, user names, audit event names and settings remain visible to anyone with file access. Use BitLocker too.
- Lost credentials: if every password and the recovery key are lost, data cannot be recovered.
- Whisper/LLM make mistakes, especially with Tagalog/Taglish, accents, noise and drug names (observed in testing: "amoxicillin" heard as "a Moxicillin"). The app flags likely mis-heard drug names but cannot guarantee correctness.
- Llama 3.2 / Phi-3.5 are not officially tuned for Tagalog; Taglish notes carry an explicit "unverified translation" marker.
- Medication data are synthetic until you import and approve verified content (
docs/REFERENCE_DATA.md). - Single workstation, no TLS, no multi-site sync. Not tested on macOS/Linux. No speaker diarisation.
- Compliance with the Data Privacy Act of 2012 requires your own privacy impact assessment, policies, DPO review and security hardening.