Disaster procedures
Plain, actionable guidance for earthquakes, fires, floods, evacuation, shelter, water safety and basic logistics.
CrisisSentinelEdge
Offline. Private. Ready in the field.
Crisis Sentinel Edge is the downloadable local model layer for disaster guidance, first-aid boundaries, field report drafting and multilingual direction. The app stays lightweight; the model is delivered later through a protected release pipeline.
0
Signal required
Local
Answer generation
Protected
Model delivery
09:41LTE
Crisis Sentinel Edge
Local release verified
Model package
CS Edge 1.0.0
Release
Local release verified
Field question
After an earthquake there is a gas smell at an apartment entrance and two people are panicking. What should we do?
Generated on-device
- If gas is suspected, do not use switches, lighters or electrical controls.
- Ventilate if safe, move people to a safe distance and notify official responders.
SITREP and task note drafting
DraftFor field teams, draft a short SITREP: location, hazard, people affected and urgent needs.
No cloud round-tripLocal knowledge baseChecksum verified
Security
Designed for institutional confidence in crisis environments
This is not a certification or endorsement claim. The product language is designed around an auditable, human-centered AI support layer that avoids adding harm during disaster response.
Release control
Model id, platform, file name and version metadata are managed through the manifest for controlled releases.
Checksum and size validation
The mobile app validates expectedBytes and SHA-256 before accepting the downloaded file.
No public model URL
The Storage path is not readable through public rules; model download goes through an App Check-enforced callable.
Local data processing
Questions and answer generation happen on-device unless the user explicitly shares operational context.
Capabilities
Built for the questions people actually ask in the field
Edge is not just a chat box. It combines actionable answers, safety boundaries, short field reporting and a knowledge architecture that still works when connectivity is gone.
First-aid support with professional boundaries
Guidance for bleeding, burns, suspected fractures, smoke exposure, hypothermia and stress reactions without pretending to replace medical care.
SITREP and task note drafting
Field team mode turns hazards, casualty counts, location hints, needs and missing details into a concise incident draft.
Multilingual crisis guidance
Designed for clearer explanations, short warnings and translation-assisted field communication in the user's language.
Ready for photo, OCR, file and voice input
The mobile connector layer is being staged for damage photos, document OCR, supply classification, sign reading and voice dictation.
Mesh and device diagnostics
Explains Bluetooth mesh status, nearby-device communication issues and offline operation in language users can act on.
Architecture
The model comes to the phone; data does not need to leave it
Crisis Sentinel Edge uses a separate release pipeline so the store build stays small and the model file never becomes an uncontrolled public link.
- 01
No model bundled in the APK or IPA
The app stays lightweight. Users download the model package in-app while Firebase manages release state.
- 02
Manifest is verified
The mobile app reads release id, file name, byte size, SHA-256 and minimum free-space requirements.
- 03
Download is App Check protected
The model is not opened through public Storage rules; verified apps request a short-lived signed URL.
- 04
On-device inference
After the file is validated, answer generation runs on the mobile runtime.
- 05
Local knowledge and safety layer
Disaster guide content, Crisis Connect knowledge, medical boundaries and safety filters review model output.
- 06
Optional sync
When a field team reconnects, draft reports can move to the institutional panel; public use can remain local guidance.
Medical and psychological safety
First-aid information with professional boundaries
Edge is not trying to be a doctor, emergency service or rescue commander. It exists to provide the safest understandable next step when the network is gone.
Priority in life-threatening risk
No breathing, unconsciousness, severe bleeding, chest pain, stroke signs, major burns, entrapment, gas or fire risk should prioritize the local emergency number and authorized responders.
Medical safety boundary
A clear safety boundary for first aid
Crisis Sentinel Edge simplifies first-aid information; it does not produce medical diagnosis, treatment or clinical decisions. It shows a safer first step while keeping professional help at the center.
Human-in-the-loop
It does not replace human decision-making; it supports field judgment
In line with the ICRC and WHO direction, Edge does not take over command structure or field judgment. It steps back under uncertainty, states risk plainly and routes users to authorized teams.
Medical and psychological safety
First-aid information with professional boundaries
Edge is not trying to be a doctor, emergency service or rescue commander. It exists to provide the safest understandable next step when the network is gone.
Standards and trust
NIST AI RMF approach
The release process documents intended use, limitations, evaluation scope, harmful-advice testing and user-facing safety boundaries.
FEMA NIMS / ICS logic
Incident management and SITREP language
Answers are shaped toward fields operations teams can process: incident type, hazard, location, people affected, needs, resources and next action.
WHO and IFRC boundaries
First aid and psychological first support
Instead of claiming diagnosis or therapy, Edge prioritizes safe first-aid guidance, emergency-service escalation, calming communication and culturally aware support.
NIST AI RMF
Measurable risk management
The release process documents intended use, limitations, evaluation scope, harmful-advice testing and user-facing safety boundaries.
OWASP LLM Top 10
LLM security threats
Prompt injection, sensitive disclosure, model supply chain, unsafe output and overreliance are handled across mobile and backend layers.
Model card preview
The open-weight plan should start with technical documentation
Before opening the model to the community, the first serious step is a clear model card that lets everyone see the intended use, boundaries and quality bar.
Model type
On-device, downloadable local language model focused on disaster guidance
Distribution target
Secure in-app model release pipeline for Android and iOS
Base model note
Gemma-based derivatives preserve Gemma Terms and NOTICE requirements
Human oversight
Does not replace field or medical decisions; escalates uncertainty to authorized teams
Evaluation matrix
Model quality is measured by release gates
For Crisis Sentinel Edge, the professional path is not only showing good answers; each release should state the domains measured, release gates and device performance thresholds.
Disaster procedure accuracy
Alignment with authorized guidance across earthquake, fire, flood, evacuation, shelter, water safety and logistics scenarios.
Answers that put steps in unsafe order, minimize hazards or conflict with official warnings do not ship.
First-aid safety boundary
Avoiding diagnosis or treatment claims for bleeding, burns, suspected fractures, smoke exposure and hypothermia.
Medication dosage, diagnosis, treatment plans or emergency-service replacement advice blocks release.
Psychological first-support language
Calm, brief, culturally aware and action-oriented wording for panic, shock, children, older adults and crowded scenes.
Blaming, panic-amplifying, dismissive or therapy-claiming answers fail release.
Multilingual translation
Short crisis warnings translated across Turkish, English and target local languages without losing emergency meaning.
Translations that reverse, weaken or obscure urgent action do not ship.
Hallucination and harmful-advice refusal
Attempts to trigger fabricated agencies, fake procedures, dangerous interventions, prompt injection and sensitive disclosure.
If the model invents certainty or cannot pull dangerous requests back into a safe boundary, release is blocked.
Offline latency / RAM / battery
Answer latency, memory use, model load time, low-battery behavior and timeout handling for long answers.
The model does not become default until target device classes meet usable latency and memory thresholds.
Open-weight roadmap
The community release ships with governance
Opening Crisis Sentinel Edge to the community makes sense, but it should not mean only publishing a model file. Evaluation, data documentation, safety policy, model card and license notices should be part of the same release package.
For a Gemma-based derivative, the cleaner language is governed open-weight or community release rather than a classic OSI open-source claim. A release should include a model card, NOTICE file, use restrictions, safety evaluations and clear references to the base model terms.
01
Community eval pack
Disaster procedure, first-aid boundary, psychological support language, translation and harmful-advice refusal tests are published reproducibly.
02
Dataset documentation
Data sources, excluded content, local language coverage, sensitive-data cleanup and medical review status are documented.
03
Safety policy
Boundaries for harmful medical advice, official authority impersonation, crisis manipulation, sensitive data and prompt injection are stated.
04
Model card
Intended use, out-of-scope use, known limitations, evaluation results, device targets and human oversight are collected in one document.
05
LoRA adapter / fine-tuned weights
If license and base-model terms allow it, the adapter or fine-tuned artifact ships with controlled release notes.
06
NOTICE.txt
Gemma Terms reference, modified model notice, release date, checksum and distribution terms ship with the model package.
Download disaster guidance before connectivity fails
Crisis Sentinel Edge lives inside the Crisis Connect mobile app: download the model, verify it and keep basic guidance available when the field phone goes offline.
Frequently asked questions
Clear answers about Edge
The most important points about how the model runs on phones, how it is secured and how an open-weight release could work.
Does Crisis Sentinel Edge need internet?
After the model and knowledge package are downloaded, core question-answering and guidance work offline. Connectivity is required for new model releases or report synchronization.
Why is the model not bundled in the APK or IPA?
Keeping the model separate avoids bloating the store build, lets releases be versioned and keeps model delivery under security controls.
Will photo, file and voice support be available?
Yes. The media layer described here is being staged for OCR, file text extraction, photo observations and voice dictation inside the Edge flow.