This chapter is PIF AI’s commitment to the future: what will be done when, how it is deployed, why AGPL-3.0, and how community contributions are accepted. After reading, you should be able to judge: is this a project you can invest in long-term as a contributor or investor?
Focus: validate that AI produces usable PIF drafts.
Delivered:
Not yet: electronic signature, batch multi-brand mode, billing, advanced ECHA, regulatory change alerts.
Focus: complete 16 items + live production + commercialization.
gantt
title PIF AI Three-Phase Roadmap
dateFormat YYYY-MM-DD
axisFormat %Y/%m
section Phase 1 MVP
Core CRUD + AI MVP :done, p1a, 2026-01-01, 2026-03-31
i18n 5 locales :done, p1b, 2026-04-15, 2026-04-20
Central RAG (backend) :done, p1c, 2026-04-19, 2026-04-22
Closed beta :active, p1d, 2026-04-20, 2026-06-30
section Phase 2 GA
E-signature + full 16 :p2a, 2026-07-01, 2026-09-30
Subscription billing :p2b, 2026-08-01, 2026-10-15
ECHA integration :p2c, 2026-09-01, 2026-11-30
GA launch :crit, p2d, 2026-12-01, 2026-12-31
section Phase 3 International
Multi-jurisdiction :p3a, 2027-01-01, 2027-12-31
Multi-region deploy :p3b, 2027-03-01, 2027-09-30
Figure 12.1: Phase 1 closed beta runs from 2026-04-20 through just before the regulatory deadline (2026-07-01). Phase 2 drives GA in H2 2026, targeting 200 paying customers by year-end (CLAUDE.md business goal).
MVP deploys on a single host (Hetzner / DigitalOcean / on-prem):
host
├── pif-frontend-1 :3000
├── pif-backend-1 :8000
├── pif-worker-1
├── pif-db-1 :5432 (pgvector/pgvector:pg16)
├── pif-redis-1 :6379
└── pif-minio-1 :9000 (S3-compatible)
External reverse proxy via Nginx Proxy Manager (NPM) handles TLS certificates, custom domains, basic WAF.
Pros:
docker compose up -d)Cons:
When MAU > 10,000 or req/sec > 100:
flowchart LR
CF[Cloudflare CDN]
subgraph K8s["Kubernetes Cluster"]
subgraph NS1["pif-prod namespace"]
FE[Frontend<br/>Deployment x3]
BE[Backend<br/>Deployment x3<br/>HPA CPU]
WK[Worker<br/>Deployment x5<br/>KEDA queue]
end
subgraph NS2["pif-data namespace"]
PG[(PostgreSQL<br/>Operator)]
RD[(Redis<br/>Cluster)]
end
end
S3[(AWS S3 / R2)]
ANT[Anthropic API]
RAG[rag.baiyuan.io]
CF --> FE --> BE --> PG
BE --> RD
BE --> S3
BE --> ANT
BE --> RAG
WK --> PG
WK --> ANT
Figure 12.2: K8s deployment uses a PostgreSQL Operator (e.g., CloudNativePG) for primary/replica; KEDA scales Workers by queue depth. External dependencies go through an egress gateway.
Phase 3 global expansion:
Each region: independent cluster + DB; DNS-based geo routing via Cloudflare Load Balancer.
| License | Closed-source derivative | SaaS loophole | Business-friendliness | PIF fit |
|---|---|---|---|---|
| MIT / Apache-2.0 | ✅ Allowed | ✅ Allowed (no disclosure) | Most | ❌ |
| GPL-3.0 | ❌ Must open | ✅ Allowed (SaaS not distribution) | Medium | ❌ |
| AGPL-3.0 | ❌ Must open | ❌ SaaS must open too | Stricter | ✅ Chosen |
| Proprietary | — | — | Closed | ❌ |
Core consideration: prevent forks from packaging PIF AI as a competing SaaS without contributing back.
MIT permits: Company A forks PIF AI → wraps as cloud service → charges for access → no disclosure of changes.
AGPL Article 13: if users access a modified version over a network, the source must be available. This forces fork-and-sell to open-source changes.
For third parties building on PIF AI:
| Scenario | AGPL impact |
|---|---|
| Brand owner uses pif.baiyuan.io | None — users are not “distributors” |
| Business self-hosts PIF AI for internal use | None (no external service) |
| Company packages PIF AI as a product for sale | Must disclose modifications |
| External system integrates with PIF AI | Depends on integration depth (see AGPL FAQ) |
For the vast majority of users (businesses), AGPL is transparent. Only fork-and-resell scenarios are affected.
The whitepaper is licensed CC BY-NC 4.0 (attribution, non-commercial), allowing:
This lets the whitepaper flow freely in academic circles while preserving licensing flexibility for commercial contexts (paid training, publications).
| Role | Responsibility | Count |
|---|---|---|
| BDFL (Benevolent Dictator For Life) | Project vision, final decisions, constitution keeper | 1 (author) |
| Maintainer | PR review, issue triage, releases | 3–5 (to be recruited before Phase 2) |
| Contributor | Submit PRs, translations, docs | Unlimited |
| User | Report bugs, participate | Unlimited |
As a Claude Code engineering case study, we welcome the Anthropic community to:
PIF AI aims to prove: LLM-assisted engineering can deliver commercial-grade open-source SaaS in the regulatory-compliance domain. This whitepaper is the full record of that process.
If you are:
Thank you for reading this far. The remaining 4 appendices provide complete reference material:
| Version | Date | Summary |
|---|---|---|
| v0.1 | 2026-04-19 | First draft. Three-phase roadmap, deployment evolution, AGPL rationale, community governance |
© 2026 Baiyuan Tech. Licensed under CC BY-NC 4.0.
Nav ← Chapter 11: Security Model · Chapter 13: Compliance Engine Deep Dive →