I build AI systems
end to end —
evidence first.

The problem

AI news arrives faster than anyone can triage it, and most of it is duplicate coverage, rumour or commentary without material change. The hard part isn't retrieval — it's high-precision triage at volume while keeping provenance intact and every decision explainable.

How it's built

• Ten-stage pipeline: discovery → normalization → source resolution → selective extraction → event candidates → semantic adjudication → publication
• Event-centric data model keeping SourceArticle ≠ Event ≠ Signal and evidence ≠ judgment ≠ decision as separate concepts
• Versioned offline evaluation harness plus blind human labeling for calibration
• Daily worker with an atomic day-claim, leases and provider budgets — a duplicate run costs nothing

The problem

Small businesses want an agent answering their DMs, not an AI project. That means real multi-tenancy, a human who can take over mid-conversation, and an agent that never invents a price or an availability.

Engineering decisions

• Every tenant-owned row carries organization_id, enforced by database row-level security
• A conversation is exactly one of AI_ACTIVE / HUMAN_ACTIVE / PAUSED / CLOSED, re-read immediately before every automated reply
• Structured business facts are queried from structured data; RAG only supplements them
• Each provider call records model, tokens, an immutable rate snapshot and the credits charged
• Provider identifiers never become the domain's primary identifiers

The problem

A model writing a study guide from memory produces something plausible and occasionally wrong — which in study material is worse than useless, because you memorise the error.

How it's built

• The generator writes only from a supplied extract; a validator fails any guide whose sources are missing
• Web-sourced mode searches, reads and builds an extract first — same pipeline, same fidelity rule
• Bilingual by design: intuition in your language, technical vocabulary in the language of the field
• Domain profiles are prompt files, not code — adding a field means copying a markdown file
• Containerised because the render step needs a real runtime: Chromium for the PDF and a writable disk

The problem

Teachers don't want a chat transcript — they want the three artefacts a class actually needs, formatted and ready to print, in under two minutes.

How it's built

• Model output is forced into a JSON lesson schema through tool choice, which removes formatting failures entirely
• Uploads reach the model as native document and image blocks — no local parsing layer to break
• Model routing by plan: a fast tier for the free plan, a stronger tier for Pro
• Generation takes 30–90s, so it runs on a Docker/VPS deploy instead of a serverless timeout
• .pptx built with pptxgenjs, PDFs with jsPDF, plus a demo mode that works with no API key

The interesting bit

The model doesn't generate an image — it acts as art director and returns pure HTML/CSS with editorial typography, rendered at full 1080×1350 or 1080×1920 and exported to PNG in the browser. Type stays crisp and the output avoids the "AI look" entirely.

How it's built

• A visual planner decides whether a theme is stronger with real photography or stays typographic
• Photography comes from Pexels with authorship and origin preserved in the history
• A curated font list guarantees the preview and the exported PNG are identical
• Plans, credits and quotas enforced in Postgres; prompt caching on the system prompt to cut cost

What it covers

Public marketing and lead qualification, consultations and onboarding, memberships and session management, then coach programming with RIR and RPE tracking, adherence, check-ins and form review.

Decisions so far

• Localized en and pt-BR routes with English fallback, on brand tokens and corrected SEO foundations
• Versioned Supabase schema with a row-level-security baseline from day one
• AI assistance runs server-side and is gated behind explicit coach approval — never auto-sent to a client
• Production DNS, billing and destructive migrations require owner approval by policy

Where it stands

Pre-launch and pre-revenue, with a complete bilingual multipage site built on implementation-first messaging, kept as a protected draft with no production deployment while the brand and pricing decisions stay open.

Why it's here

It's the clearest example of how I run a project: every consequential decision is written as an accepted decision record, status is reported as it is rather than as it's hoped, and nothing ships to a public domain before the gate is genuinely met.