MarketEQ – Platform Architecture

LAYER 1

Data sources

Curated sources around the Swiss insurance market. Every source is continuously measured by its hit rate (source precision) — weak feeds are dropped.

Insurer newsrooms

Press releases from primary insurers & reinsurersRSS-Feeds

Brave Search API

Targeted web searches per insurer & vendorREST · 12 search profiles

Trade press & events

Industry media, conference materialRSS / Sitemaps

LinkedIn & tips

Manual finds — submitted via Telegram linkvia Telegram bot

Artificial Analysis

Benchmark data for all major AI modelsREST API · daily

Annual reports

KPIs of the 15+ largest EU insurers, CHF-normalisedcurated + sourced

automated crawls (weekly + daily) · manual submission (instant)

LAYER 2

Automation layer

Two orchestration services share the work: scheduled crawler jobs and the interactive review/intake workflow.

T.Dcloud.trigger.dev — scheduled jobs

discover-sources · Sun 10:00 — finds new solution signals
process-leads · every 6 h — relevance gate → extraction → draft
refresh-existing-sources · Sun — new sources for existing entries
refresh-ai-models · daily — refresh AI model benchmarks
saved-solution-digest · daily — email on changes to watched solutions

n8n Cloud — interactive agent

AI review · daily 07:00 or instantly via the Telegram command /review
Telegram intake: send a link to the bot → the source is read, fields extracted, an entry created or routed as an update
Replies come back as a Telegram message (incl. publication links)

Telegram Bot API ↔ n8n ↔ Supabase REST ↔ OpenAI

EVERY signal — whether from the crawler or a human — passes the same gates

LAYER 3

AI quality gates

The core — the USP

Why it matters: Generic AI directories collect marketing claims. MarketEQ publishes only what a named insurer demonstrably deploys — every data point with a graded source, every approval cross-read against the live source.

Relevance gateOnly concrete deployments at named (re)insurers, Swiss market scope. Threshold calibrated from data (0.75). Learns from moderator rejections. OpenAI gpt-4o-mini

Extraction & source ratingStructured fields in EN + DE, enum-validated; every source graded by the Admiralty Code (reliability A–F, credibility 1–6). Nothing is invented — unsourced fields stay empty. OpenAI gpt-4o / gpt-4.1

AI review (5 criteria)Cross-reads the primary source live against: relevance, deployer vs. vendor, DE/EN completeness, source support, EU AI Act plausibility. Approval only at ≥ 75 % confidence + source backing. OpenAI gpt-4.1

Duplicate gate (3-stage)Before every auto-publication: ① exact source URL ② fuzzy name match (pg_trgm) ③ semantic AI judgement against all entries of the same insurer. When in doubt: stop → a human decides.Postgres pg_trgm + gpt-4.1

↺ Reject-feedback loop: Every moderator rejection reason is stored (moderation_feedback) and flows back automatically as a negative example into the filter prompts — the gates get stricter with each human decision.

passed → auto-published · doubt → escalated to the moderator · irrelevant → discarded with a reason

LAYER 4

Supabase — database, auth & row-level security

Postgres as the single source of truth. Access is secured by role (RLS): only status = published is publicly visible. Example tables:

solutionsname · name_de · description(_de) · provider (insurer) · implementation_partner (vendor) · sources jsonb[url, reliability A–F, credibility 1–6] · status · eu_ai_act_risk…The catalogue — one entry = one deployment

crawl_logsource_url · relevance_score · lead_state · crawl_status · discovery_channelEvery signal, auditable incl. reason for discard

source_update_candidatessolution_id · url · is_material · ai_diff_summary · suggested_reliabilityNew sources for existing entries (append-only)

moderation_feedbackkind · reason · context · created_byLearning loop: rejection reasons → prompts

profiles / rolesrole: viewer · plus · professional · team · moderator …Feature gating per subscription tier

ai_modelsintelligence_index · pricing · speed · benchmarksDaily snapshot from Artificial Analysis

watchlists + notification_prefsuser_id · solution_id · digest opt-outBasis for the daily email digest

solution_confidence_assessmentsfield_name · confidence_level · noteConfidence per data field — shown in the frontend

Server rendering with role-based access

LAYER 5

Platform — ai.marketeq.net (Next.js on Vercel)

Catalogue & detail pages

Solutions with source backing, confidence badges, DE/EN

Market Landscape

Heatmap, gap analysis, vendor landscape, insurer KPIs

Analytics & Insights

Opportunity matrix, maturity model, comparisons, ranking

AI Models Dashboard

Public model benchmarks, updated daily

Moderation cockpit

Review queue, publish log, source precision, duplicate merge

Connected services:

Stripe

Subscriptions (Free → Plus → Professional → Team); webhook assigns roles automaticallyCheckout + Webhooks · CHF

Resend

Daily email digest for watched solutions, GDPR-compliant with an unsubscribe linkEmail API

BPBotpress

EQ chatbot on the website — questions about the platform & catalogueWeb widget

Vercel Web Analytics

Cookieless visitor stats + own first-party trackingno cookies / raw IP

Operator's control channel: submitting, /review, publication confirmationsBot API

Ongoing development — even on the go

LAYER 6

Development & operations — AI-assisted

The platform itself is developed with AI: Anthropic Claude models build, test and deploy changes — traceable via Git and a Kanban board.

Claude Code (Anthropic)

Claude models build features, run audits & end-to-end tests, and deploy to Vercel + Trigger.devAnthropic Claude · Fable/Opus

Claude Dispatch

Trigger changes from your phone: describe what you want → Claude implements it → pull request → deploy. No laptop needed.mobile → PR → live

GitHub

Version control, pull requests as change docs, audit trail of all migrationsgit · gh CLI

Todoist — "Platform Tasks"