"Find the Missing Piece"
Adaptive, AI-Supported Stream Assessment with Human-in-the-Loop Integrity
"Don't ask everything. Ask what matters." AquaAI is an adaptive intelligence layer built for urban stream monitoring in the IEEE OneAquaHealth Global Hackathon 2026 (Primary: Track 3 AI-Supported Assessment; Secondary: Track 1 Citizen Science UX & Track 7 Digital Health Standards).
The OneAquaHealth Citizen Science App invites community volunteers to assess urban streams through an exhaustive, multi-field form (bank stabilization, impervious margins, water clarity, substrate composition, habitat structure). In practice:
- Volunteers skip questions, become overwhelmed, or answer with "I'm not sure".
- Observations frequently contradict attached photos or spatial context.
- Environmental scientists cannot distinguish trustworthy submissions from noise, requiring costly manual triage.
- Lengthy forms deter repeat community participation, degrading long-term One Health surveillance.
AquaAI solves this with an adaptive, human-in-the-loop intelligence layer. Instead of demanding a 20-field questionnaire upfront, it evaluates existing evidence, detects the highest-value information gap, and poses one plain-language question at a time until the observation reaches scientific readiness.
[ Citizen Observation / Photos ]
│
▼
┌───────────────────────────┐
│ Evidence Reliability │ ◄── Rule-based & fully explainable
│ Scoring Engine │ (Completeness + Consistency + Photo Agreement)
└─────────────┬─────────────┘
│
┌────────────┴────────────┐
▼ ▼
Score < 80% Score ≥ 80% (Ready)
┌───────────────────────┐ ┌──────────────────────────┐
│ Information-Gap │ │ Publish Standardized │
│ Targeted Questioning │ │ HL7 FHIR R4 Observation │
│ (Max gain per effort) │ │ (Ready / Verified Tier) │
└───────────────────────┘ └──────────────────────────┘
AquaAI bridges the gap between everyday community stream observations and rigorous ecological research:
- Adaptive, Step-by-Step Observation: Replaces intimidating 20-question checklists with a friendly, guided flow. Users start with simple free-form notes, photos, and optional GPS coordinates.
- Dynamic Evidence Completeness Ring: Computes an explainable completeness score (ranging from 60% to 100%) that updates in real time based on geographic location, photo evidence, description detail, and answer clarity.
- Targeted Information-Gap Questioning: Automatically calculates which ecological indicator provides the highest scientific value and asks just that question with plain-language choices.
- Photo-vs-Answer Conflict Detection: Cross-checks human answers against computer-vision photo signals, detecting contradictions (e.g., reporting "natural plants" when photos reveal concrete embankment).
- Human-in-the-Loop Conflict Protection: When a conflict occurs, the system gently asks the volunteer to take a second look. If the volunteer re-confirms their finding, their voice is preserved and routed to Expert review—the AI never silently overrides human input.
- Dual Citizen & Researcher Interfaces:
- Citizen View: Focused questionnaire, previous-answer editing, custom answer inputs ("Others:"), and instant printable PDF environmental reports.
- Researcher Dashboard: Live triage queue sorted by urgency (
Expert review,Needs more info,Ready), detailed audit trails, conflict inspection, and one-click verification.
- Standardized Digital Health Publishing: Exports every completed observation into an HL7 FHIR R4
Observationresource with custom extensions for evidence scores and review tiers.
AquaAI operates as an uncertainty-reduction loop composed of five coordinated stages:
[1. Ingest Evidence] ──► [2. Explainable Score] ──► [3. Detect Gap & Ask] ──► [4. Route & Triage] ──► [5. Export FHIR R4]
When a citizen submits an observation:
- Free-text descriptions are parsed by an allow-list validation layer for recognized environmental attributes (water clarity, bank materials, substrate).
- Attached photos are evaluated for visual signals (vegetation, concrete, algae, debris).
- GPS coordinates and observation metadata are registered.
The core scoring engine in backend/engine.py computes evidence completeness without black-box neural networks. Every score traces directly to weighted indicators:
- Corroborated answers (photo agrees with citizen):
1.00 - Standard answers (no contradiction):
0.85 - Re-confirmed answers under conflict:
0.60(routes to Expert Review) - Uncertain / Ambiguous answers:
0.25 - Unresolved contradictions:
0.20 - Missing indicators:
0.00
Criteria adjustments apply based on GPS verification (+5%), photo corroboration (+7%), narrative detail (+3%), and consistency warning flags (-3% each).
The engine scans all remaining unobserved indicators and scores them using the expected gain per unit of volunteer effort:
Every record is classified into an actionable review tier:
-
Ready (
$\ge 80%$ ): High confidence, complete evidence, and consistent signals. Ready for immediate ingestion into environmental dashboards. - Usable – verify (70% – 79%): Sufficient for general trends; optional quick scientist sign-off.
-
Needs more info (
$< 70%$ ): Sparse details; prompts citizen for an additional photo or clarification. - Expert review (Any score with contradiction): Hard conflict between photo signal and citizen selection. Preserved for limnologist manual inspection.
Outputs valid HL7 FHIR R4 JSON documents carrying answers as component slices and metadata as extension fields, bridging community science into institutional digital health systems.
-
Step 1: Observe (
/)- Enter what you notice at the stream (e.g., "The water looked clear, with plants on both sides").
- Click Choose photos to upload one or more stream pictures (with instant thumbnail preview and remove buttons).
- Optionally toggle Use my location to attach GPS coordinates with visual status confirmation.
- Click Next: Review.
-
Step 2: Review Initial Clues
- AquaAI extracts recognized attributes and displays them as gentle suggestion chips.
- Click Confirm to accept an extracted clue, or leave it for manual question answering.
- Click Next: Answer questions.
-
Step 3: Answer Targeted Questions
- AquaAI presents only the single most impactful question right above the answer choices.
- Select an option, or choose "Others:" to type in a custom answer.
- Made a mistake? Click Previous or jump directly via the step navigation bar (
1 Observe,2 Review,3 Answer,4 Done) to revisit and edit earlier responses. - If a photo conflict is detected, AquaAI highlights the discrepancy and allows you to either adjust your answer or re-confirm what you personally witnessed.
-
Step 4: Done & Export
- View your dynamic Evidence Completeness Ring (60% to 100%).
- Expand How AquaAI decided to inspect the exact mathematical gain and runner-up questions.
- Click Download Summary Report to generate a clean, printable PDF report containing observation details, review tiers, and complete audit history.
- Accessing the Dashboard:
- Click the Researcher · demo access tab in the top navigation bar.
- Reviewing the Triage Queue:
- View all incoming observations sorted into categories:
- Expert review: Flagged contradictions awaiting limnologist adjudication.
- Needs more info: Observations requiring additional volunteer follow-up.
- Ready / Verified: High-reliability submissions ready for environmental analysis.
- View all incoming observations sorted into categories:
- Auditing & Adjudication:
- Click any queue item to open its comprehensive dossier: view citizen statements, photo evidence, spatial coordinates, and full audit logs.
- Click Mark verified to approve the record, or Request more info to flag it for field re-survey.
- Exporting FHIR R4 Data:
- Inspect the live HL7 FHIR R4 Observation JSON directly within the dashboard for seamless export into environmental GIS and public health databases.
- Validation Checks (Photo-vs-Answer Consistency): Detects contradictions between user selections and photo-derived signals (e.g., reporting "natural vegetated banks" while photos indicate concrete channel walls).
- Explainable AI (XAI): Every single percentage point traces directly to a specific field weight, consistency cross-check, or verified detail. No black-box opacity.
- Human-in-the-Loop (HITL) Workflow: Contradictions are routed to human expert review—never silently altered or overridden by AI. If the citizen re-confirms their finding against the AI's suggestion, the citizen's word is preserved and prioritized for specialist inspection.
- Zero Questionnaire Fatigue: Transforms a tedious multi-page checklist into a guided 3-minute conversation.
- Dynamic Gain-per-Question: Prioritizes questions offering the highest information gain per unit of cognitive effort.
- Supportive Feedback: Inline guidance explains why the question is asked and how it advances stream assessment.
- Emits standard HL7 FHIR R4
Observationpayloads. - Includes citizen answers as structured
componentelements, along with custom extensions for the reliability score and review tier. - Aligns with the OneAquaHealth FHIR Implementation Guide to unify citizen science reporting with sensor arrays and laboratory assays.
The reliability score measures how well-supported and consistent an observation is—it is explicitly not a probability of truth or an ecological health diagnosis.
Each ecological field carries an assigned importance weight. A field's contribution is computed based on its evidence state:
| Field State | Weight Factor | Description |
|---|---|---|
| Answered & photo agrees | 1.00 |
Citizen answer cross-checked and corroborated by photographic signals. |
| Answered, unverified | 0.85 |
Citizen answer recorded without contradictory evidence. |
| Contradiction re-confirmed | 0.60 |
Citizen reviewed the photo contradiction and maintained their answer; routed to Expert review. |
| "I'm not sure" / Ambiguous | 0.25 |
Citizen indicated uncertainty or low visibility. |
| Answer contradicts photo | 0.20 |
Unresolved conflict between citizen answer and image evidence; lowers score and triggers triage. |
| Missing | 0.00 |
Unanswered field. |
The engine selects the next question by identifying the unanswered field that maximizes expected information gain per unit of volunteer effort:
Questions falling below a minimum gain threshold are omitted entirely, ensuring volunteers are never over-questioned.
Benchmarked using the automated test suite and scripted scenario evaluations in eval.py:
- Contradiction Resolution in a Single Step: An observation with an unresolved conflict between "natural banks" and photo evidence of concrete walls started at 77.4% (Expert Review). After a single targeted prompt asking the citizen to re-check the bank substrate, the score rose to 89.6% (Ready).
-
Rapid Convergence without Full Forms: Starting from a sparse initial note at 37.7%, the engine selected questions in descending value order:
$$\mathbf{37.7%} ;\longrightarrow; 46.4% ;\longrightarrow; 55.0% ;\longrightarrow; 62.2% ;\longrightarrow; 68.0% ;\longrightarrow; 75.2% ;\longrightarrow; \mathbf{82.4%} ;\longrightarrow; \mathbf{89.6%}$$ The assessment crossed the scientific readiness threshold ($\ge 80%$ ) in just 6 questions, saving up to 70% of the input effort required by conventional survey forms. - Human-in-the-Loop Integrity: When a user insisted on an answer contradicting the computer vision signal, the system recorded the citizen's confirmation, maintained the score at 83.5%, and locked the status to Expert review for scientist verification.
┌────────────────────────────────────────────────────────┐
│ Frontend (Web UI) │
│ Vanilla Modern HTML5 / CSS Glassmorphism / Vanilla JS │
│ - Multi-photo upload with instant thumbnail preview │
│ - Live question flow with previous answer navigation │
│ - Dynamic Evidence Ring (60% - 100% variance) │
│ - Researcher review queue and one-click PDF reports │
└───────────────────────────┬────────────────────────────┘
│ HTTP / JSON API
▼
┌────────────────────────────────────────────────────────┐
│ FastAPI Application Layer │
│ (backend/main.py) │
│ Endpoints: /api/observations, /api/fields, │
│ /api/review-queue, /api/observations/{id}/fhir │
└─────────────┬────────────────────────────┬─────────────┘
▼ ▼
┌───────────────────────────┐ ┌──────────────────────────┐
│ Scoring & Gap Engine │ │ Validation & Conflict │
│ (backend/engine.py) │ │ (backend/validation.py) │
│ - Explainable weights │ │ - Photo-vs-answer rules │
│ - Gain-per-effort triage │ │ - Suspicious input flags │
│ - Evidence completeness │ │ - Deterministic checks │
└─────────────┬─────────────┘ └──────────────────────────┘
▼
┌────────────────────────────────────────────────────────┐
│ HL7 FHIR R4 Interoperability │
│ Standard Observation resource with reliability scores, │
│ component observations, and review tier extensions. │
└────────────────────────────────────────────────────────┘
- Python 3.10+
- Modern Web Browser (Chrome, Edge, Firefox, Safari)
git clone https://github.com/Eusha-Kayenat/AquaAI.git
cd AquaAI
python -m pip install -r requirements.txt# Using the launcher script:
python run.py
# Or directly with uvicorn:
python -m uvicorn backend.main:app --host 127.0.0.1 --port 8000Open your browser to: http://127.0.0.1:8000
AquaAI includes a full unit test suite and synthetic evaluation pipeline:
# Run all unit tests (15 test cases):
python -m unittest discover -s tests -v
# Run synthetic evaluation benchmark:
python eval.pyQuestion-selection agreement: 60.0% (12/20)
Conflict-detection precision: 25.0% (2/8)
Conflict-detection recall: 50.0% (2/4)
Vague-word inputs left null: 100.0% (20/20)
AquaAI transforms citizen observations into standard FHIR R4 JSON payloads ready for public health registries and environmental monitoring sandboxes:
{
"resourceType": "Observation",
"id": "obs-6410e1f9e9",
"status": "preliminary",
"code": {
"coding": [{
"system": "http://oneaquahealth.eu/fhir/cs/stream-observation",
"code": "stream-habitat-assessment",
"display": "Stream Habitat and Ecological Assessment"
}]
},
"extension": [
{
"url": "http://oneaquahealth.eu/fhir/StructureDefinition/evidence-reliability-score",
"valueDecimal": 89.6
},
{
"url": "http://oneaquahealth.eu/fhir/StructureDefinition/review-tier",
"valueString": "Ready"
}
],
"component": [
{
"code": { "text": "Water appearance" },
"valueString": "Clear"
},
{
"code": { "text": "Left bank material" },
"valueString": "Natural vegetation"
}
]
}-
Evidence Quality
$\neq$ Ecological Truth: The reliability score gauges information completeness, clarity, and consistency. It does not measure ecological water purity or diagnose biological contamination. -
AI Never Overrides the Citizen: Machine vision outputs are strictly presented as suggestions (
"AI suggestion — please confirm"). In any contradiction, the human volunteer or researcher review retains final authority. - Calibrated Field Data Needed: Current weights and thresholds reflect expert-designed defaults. The immediate next milestone is empirical calibration against real citizen-vs-limnologist datasets from OneAquaHealth pilot basins.
- Deterministic Foundation: Core scoring and routing operate entirely without generative hallucinations. Optional LLM integrations are restricted to plain-language phrasing and strictly forbidden from modifying numerical scores.
AquaAI/
├── backend/
│ ├── main.py # FastAPI application routes & static file server
│ ├── engine.py # Core scoring engine, weights, gain math & questions
│ ├── service.py # Observation store, seed data & orchestration
│ ├── validation.py # Conflict rules & data integrity checks
│ ├── ai_provider.py # Deterministic mock provider & LLM interfaces
│ └── prompts/ # Versioned, auditable system prompts
├── frontend/
│ ├── index.html # Single-page interface (Citizen + Researcher views)
│ ├── demo.js # Client-side state transitions & demo controls
│ └── favicon.png # Official AquaAI brand icon
├── tests/
│ ├── test_api.py # FastAPI HTTP integration tests
│ ├── test_core.py # Pure engine and scoring unit tests
│ └── eval_set.json # Synthetic benchmark scenarios
├── eval.py # Synthetic evaluation script
├── run.py # Standalone application entrypoint
└── requirements.txt # Python dependencies
Developed for the IEEE OneAquaHealth Global Hackathon 2026.
Built in compliance with the OneAquaHealth One Health Framework for urban aquatic ecosystem preservation.
