System Overview

OpenScouter is a multi-layer platform that connects neurodivergent testers with businesses commissioning accessibility studies. The architecture centers on a 4-agent AI pipeline that processes live test data, applies corroboration logic, and generates structured reports. A human-in-the-loop gate ensures no AI-generated finding reaches a report without tester confirmation.

Platform Components

OpenScouter consists of five primary components that work together across each study lifecycle.

Component	Technology	Role
Web Application	Next.js 14, React	Business and tester dashboards, study management
Chrome Extension	Manifest V3	Live data capture during tests
API Layer	Next.js API routes, Supabase	Data persistence, webhook handling, agent orchestration
AI Pipeline	OpenAI, Anthropic, Google Gemini	4-agent sequential analysis
Database	Supabase (PostgreSQL)	Structured storage with Row-Level Security

The 4-Agent AI Pipeline

Each test moves through four agents in sequence. Agents operate on data produced by earlier stages. The pipeline is designed so failures in one stage degrade gracefully rather than blocking the entire study.

Agent 1: Scouty

Endpoint: POST /api/ai

Scouty handles real-time chat and initial analysis during a live test. It ingests the three parallel data streams from the Chrome extension as events arrive: browser interactions, facial expressions, and voice transcripts. Scouty identifies candidate barrier moments by correlating signals across all three streams and maintains a running list of moments with preliminary severity flags.

Scouty is the only agent that operates in real time during the test. All other agents process data after the test concludes.

Agent 2: Analyst

Endpoint: POST /api/sessions/[id]/analyze

The Analyst runs immediately after a test ends. It takes Scouty’s candidate list and performs deep-dive analysis on each moment. The Analyst maps every barrier to the relevant WCAG success criterion, assigns a severity rating, and identifies cross-task patterns where the same barrier surfaces multiple times.

The Analyst produces the structured finding objects that the tester reviews in the confirmation interface. Its output is stored as session_notes records with a confirmed flag set to false.

Human-in-the-Loop Gate

After Agent 2 completes, the pipeline pauses. The tester reviews every AI-generated finding, confirming, rejecting, or annotating each one. Agent 3 does not run until the tester sets notes_confirmed = true on the study.

Caution

Agent 3 is gated on notes_confirmed = true. Calling POST /api/reports before the tester confirms findings will return a 409 Conflict response.

This gate is not optional and cannot be bypassed via the API. It is a core product invariant that keeps tester expertise authoritative over AI output.

Agent 3: Report Writer

Endpoint: POST /api/reports

Trigger condition: notes_confirmed = true on the parent study

The Report Writer transforms confirmed findings into the individual tester report. It generates both technical descriptions and Plain English summaries for every finding. It structures output according to the selected tone profile and produces WCAG references with success criterion codes and conformance levels.

The Report Writer creates one reports record per tester per study.

Agent 4: Synthesizer

Endpoint: POST /api/reports/job/[id]

The Synthesizer runs after all tester reports for a study are complete. It performs cross-test analysis, comparing findings across all testers to identify which barriers are widespread versus profile-specific. It produces the ND stratification data that appears in the final business-facing study report and stores output as a job_reports record.

Model Strategy

The pipeline uses multiple AI providers: OpenAI, Anthropic (Claude), and Google Gemini. The platform implements automatic failover across providers. When the primary provider is unavailable or returns an error, requests are routed to the next available provider.

Every AI response includes a usedFallback boolean field. This field is stored on the relevant database record so monitoring queries can track fallback rates over time. A sustained increase in usedFallback = true records indicates primary provider degradation.

Tone Profiles

The Report Writer applies a tone profile when generating report text. Tone profiles affect word choice and framing but do not alter the factual content of findings.

Profile	Description
Standard	Direct, neutral technical language. Suitable for developer-facing reports.
Supportive	Encouraging framing. Acknowledges tester effort and product progress.
Moderate	Balanced tone between Standard and Supportive. General-purpose default.
Restorative	Trauma-informed language for studies involving sensitive user groups.

The tone profile is set at the study level when the business creates the study. It cannot be changed after the Report Writer has run.

Data Flow Summary

The following stages describe how data moves from a live test to a delivered report.

1. Tester installs the Chrome extension and begins a test
2. Extension streams browser events, facial snapshots, and voice transcripts to the API
3. Scouty processes incoming events and maintains candidate findings in real time
4. Test ends; Analyst runs and produces structured session_notes records
5. Tester reviews findings in the confirmation interface and sets notes_confirmed = true
6. Report Writer runs and creates a reports record for the tester
7. When all testers for a study have confirmed, the Synthesizer runs and creates a job_reports record
8. Business receives the completed study report

Technology Decisions

Next.js API routes handle all server-side logic. This keeps the API collocated with the frontend without requiring a separate server process.

Supabase provides the PostgreSQL database, authentication, Row-Level Security, and real-time subscriptions. RLS policies enforce data isolation between organizations at the database level, not just the application level.

Manifest V3 is used for the Chrome extension because it is the current and supported extension platform. Service workers replace background pages, which required architectural adjustments to how the extension manages state between events.