ZenAlphaLab — an agentic research platform, end to end

The problem

Investment research is repetitive analytical work: pull the same data, apply the same scoring rules, write the same structured summary, publish. The platform was built to test how much of that loop an AI agent could own end to end — not as a demo, but as a system serving real users with real money attached.

Approach

• 01The delivery platform came first: Next.js 16 App Router, NextAuth v5 with JWT sessions, Prisma 7 against Supabase Postgres, and Stripe subscriptions with webhook-driven role changes across four access tiers.
• 02A deterministic scoring layer turns live quotes into a signal — trend versus 50/200-day moving averages, valuation against peers, volume confirmation, and contrarian checks for overbought or oversold extremes.
• 03An agentic workflow wraps it: pull authoritative quote data, gather news context, apply the scoring rules, write structured JSON, trigger a Cloud Build deploy. A content refresh is one command rather than an afternoon.
• 04The human stays in the loop deliberately — the agent proposes signals and flags its own uncertainty, and every published change is reviewed before it ships.

Result

• +Content refresh went from a manual multi-hour task to a single command with a ~3.5 minute build-and-deploy.
• +The scoring layer is explicit and auditable — every signal traces to named thresholds rather than a model's opinion.
• +Runs on Cloud Run with secrets in Secret Manager, bilingual EN/ZH content, and role-gated access enforced in middleware.

Stack