Francisco Acasuso · AI Strategist & Builder

Saga: the week of drafts written in your voice, and the ways to start a post. Nothing posts without approval.

Category: Client SaaS (Stratas product, Stripe-billed) Status: Live, active development Time window: May 2026 → ongoing (380 commits since split from monorepo) Live URL: https://saga.stratas.no (legacy alias saga.stratas.no)

Purpose

A LinkedIn content copilot that drafts weekly posts from a user's own work, schedules them, and posts to personal profiles or company Pages – with a two-sided confidentiality gate that stops client pipeline data from ever entering the prompt.

Who it is for

Operators, founders, and consultants who want to publish on LinkedIn from real work but do not have hours per week to draft and schedule.

Stack

Frontend: Next.js 15 (App Router) + React 19
Auth: NextAuth (LinkedIn OAuth, dual-app flow for personal login + org posting)
Backend: Trigger.dev v4 (per-user weekly schedules.task, 3-retry budget)
Data: Supabase (dedicated linkedin schema)
LLM / AI: Anthropic Claude – Opus 4.6 (chat), Sonnet 4.6 (batch drafts), Haiku 4.5 (verification only)
Payments: Stripe
Observability: Langfuse (every Claude call traced with resolved user ID)

My contribution

Shaped the copilot through direct product use, investigated the confidentiality incident and translated it into product and engineering requirements. Designed and implemented input and output checks, added regression tests, and carried the changes into production and ongoing use.

Notable capabilities

Two-sided confidentiality gate: detectConfidentialLeak on the way out + isConfidentialMetric + protectedClientNames() on the way in
Model stratification with cost tags per call; per-route/context spend totals in Langfuse
LinkedIn scope parsing tolerant of both comma and space delimiters (LinkedIn returns granted scopes both ways)
Per-user weekly draft scheduling migrated from Vercel cron (global) to Trigger.dev v4 (per-user, retryable)
Vision-pass reads of user dashboards without leaking dashboard UI vocabulary as proper nouns

Outcome

208 test files with 3,485+ assertions. Three eval suites (brand adherence, draft quality, framing A/B, avatar likeness). Confidentiality regression that triggered PR #345 has not recurred. 80%+ coverage house rule enforced.

Evidence

Case study: Related case study
Live: saga.stratas.no
Key files: shared/publish-gate.ts, shared/work-prompt.ts, src/lib/linkedin-org-oauth.ts, src/lib/claude.ts, evals/*.eval.test.ts

Skills demonstrated

nextjs-15, react-19, nextauth-linkedin-oauth, trigger-dev-v4, supabase-schema-isolation, claude-opus, claude-sonnet, claude-haiku, langfuse-per-call-tagging, stripe, ai-safety-gates, evaluation-suites, oauth-scope-parsing, vitest

Saga – LinkedIn content copilot