Paragon Full-Stack Software Development

Current State

The project originates in the 19 June 2026 discussion on no-code versus owned full-stack. As of the 24 June design session it is in an R&D / prototyping phase within the existing budget — no new fees or subscriptions.

The working direction is firming — database-centric, AI-central, reduced role rigidity, a questionnaire-centric document model (see below) — but no formal decision has been made. The plan is to see what the prototypes look like over the next month or two, then decide whether to continue. The current no-code portal stays in production throughout.

Architecture Direction

Stack

Next.js (front end) + Supabase (database) + Vercel (deployment). Code in GitHub, owned by Paragon (no data stored in GitHub). TypeScript throughout.

AI & PII handling

Google Vertex AI as the PII-safe "walled garden" (pending confirmation it is GDPR-sufficient and multimodal). AI targeted at specific steps, not an AI-orchestrated system — most workflows are deterministic automations, with agents only on the periphery.

Build method

Spec-driven AI development: architectural decisions and specs first, then agent execution, with full-stack ownership of the code. Small team led by Joel.

Migration approach

Reuse the existing data model (port via Noloco's GraphQL API + Airtable export); replicate current functionality first to reach a working system fast, then reimagine selected document workflows. Tradeoff: replicate-first reaches a working system quickly; reimagining is deferred to selected workflows rather than attempted all at once.

Data & UX model

• Everything in one relational database.
• The project surfaces as "one big questionnaire" with backing documents; sections served up dynamically per stakeholder.
• Conditional logic (e.g. school/church → no PSS; toggle documents on/off).
• Wire-info form built in (no Adobe).
• Foreign-requirements check largely automated (AI checks ICNL / CSF live), team-reviewed.
• Broad document categories rather than many specific document types.

Roles

Reduce the rigid per-document uploader / reviewer / approver model — any team member can act within clear boundaries; keep history without hard-enforcing it; allow a client to assume the grantee role, and multiple grantee-side people to access the questionnaire.

AI analysis layer

A PII-protected layer (Vertex AI) over the whole project that surfaces red flags across grantee and team answers and uploaded documents, reviews governance documents and local law, and can run live local-law research.

Open Questions

• Replatform & licensable product. Should Paragon replatform off the no-code stack (Noloco / Airtable) into owned full-stack software, and is there a licensable product in what it would build? (Surfaced across three June 2026 conversations.)
• Scope & cost. Scoping in progress: total and ongoing cost versus the current stack, AI-mapping the existing Noloco / Airtable structure, the available talent pool, and the licensing angle.
• Backend approach. Build the backend in-house versus partner on a platform like Zagger (its "AI brain" model). Tradeoff: data ownership versus vendor lock-in. (Surfaced 26 June.)
• Mockup-first step. Build a mockup within current scope — no new fees or subscriptions — to assess scope, timeline, and cost before committing to a migration. (Discussed, not decided.)
• AI hosting. Google Vertex AI (existing enterprise contract) versus a self-hosted model. Tradeoff: convenience versus a stronger PII / GDPR posture. Leaning toward Vertex as the walled garden if it proves GDPR-sufficient and multimodal. Open question still to be researched: whether Vertex offers a materially different PII / GDPR posture than a Claude-based approach.
• Reimagine versus improve. Rebuild to reimagine the client / grantee experience, or simply improve the current model? Current lean (23 June): reuse the data model, replicate functionality first, audit document workflows, then reimagine selectively.

Decisions

No formal decisions yet. The items above are working direction, not commitments.