— Technical Walkthrough —

Under the hood.

For anyone who wants to understand what the platform is actually doing. No marketing. Just architecture.

§01 — The pipeline

From raw email to executive action item.

The email intelligence pipeline processes ~100 emails/day through a 3-stage cascade.

Step 1

Rule-based Prefilter

Sender domain/pattern matching. Catches noreply@, notifications, newsletters. ~22% filtered. Zero cost.

Step 2

Haiku Batch Classifier

20 emails/batch. 3-way classification: action_required / awareness / noise. ~$0.0005/email.

Step 3

Sonnet Deep Analysis

Priority scoring, action summaries, linked contacts, entity extraction. ~$0.015/email.

Step 4

Report Injection

Top 8 action items → CEO Daily. Top 5 awareness → ELT Weekly. PDF generation via Puppeteer.

§02 — Architecture

Built for durability. Observable by design.

Every component is hosted, versioned, and monitored. Zero shared infrastructure with external clients.

DATABASES

6 Supabase Projects

Reporting, RAG, MADE AI OS, Private, Transfer, Joe Private. PostgreSQL 15, RLS enforced, PITR.

FRONTEND

2 Lovable Apps

madescientificai.com (client-facing) + madeaisandbox (internal cockpit). React + TypeScript + Vite.

AGENTS

40-Agent Fleet

30 launchd crons + 10 event-driven. Node.js 22. Claude Sonnet 4 primary, Haiku 4.5 classifier.

SERVERLESS

18 Edge Functions

Supabase Edge (Deno). Brain-search, brain-router, skills-scaffold-gap, content-gen. JWT verified.

RAG

295K Chunks

Voyage voyage-3 1024d. IVFFlat index. Hybrid search (vector + FTS). 95ms p50 latency.

AUTOMATION

n8n + launchd

8 n8n workflows (VPS). 30 launchd plist jobs (local). Eyes dashboard monitors all.

AI MODELS

Claude Sonnet 4 + Haiku 4.5

Anthropic primary. Voyage for embeddings. Zero OpenAI surface. Zero-data-retention agreements.

OBSERVABILITY

Eyes Dashboard + Health Pulse

5-min launchd scan. Failure detection <10 min. Structured logs. Alert via webhook.

§03 — Automation

30 jobs keep the platform alive between human touches.

Every agent runs on a schedule. Every failure is logged. Every recovery is automatic.

30
Scheduled Jobs
5 min
Health Check Interval
<3%
24h Failure Rate
8:35 AM
Report Generation
6 AM / 2 PM
SFDC Sync
$8/day
Est. AI Cost

§04 — Security & compliance

Security & governance.

Designed for a 150-person regulated CDMO environment.

Row-Level Security

RLS enforced on all 6 databases. Per-user data isolation. No shared service keys in frontend.

Vault for Secrets

All API keys in Supabase Vault or GitHub Secrets. Zero hardcoded credentials in source.

3-Tier Confidentiality

PROJ-028: Public / Internal / Restricted classification. 30-prompt red-team test suite.

Agent Permissions Protocol

.agent/protocols/permissions.md: always/approval/never rules. No autonomous destructive writes.

§05 — Data flow

How data flows through the system.

Six databases with clear ownership. Data flows downstream, never back.

SFDC (CRM) ────────┐
Monday (OKRs) ──────┤
Gmail (Email) ──────┤──→ [AGENTS] ──→ Reporting DB ──→ CEO Daily / ELT Weekly
LinkedIn ───────────┤                      │
Firecrawl (Web) ────┘                      ├──→ RAG DB (295K chunks)
                                           ├──→ AI OS DB (skills, agents, evals)
                                           └──→ Transfer DB (competitive intel)

[LOVABLE APPS] ←── read ── Reporting DB + AI OS DB
[EDGE FUNCTIONS] ←── invoke ── Lovable Apps + n8n + launchd agents

That's how it works.

Questions? This is the actual production system — happy to walk through any layer live.