PM-DELTA
AI-Native Development Infrastructure
Shanghai Innovation Institute · 2026
From Tools to Value
PM-DELTA
Delivers business value, not just developer productivity.
Cursor, Copilot, Devin
Make individual developers faster. Single-task scope.
Claude, GPT, etc.
Provide intelligence. We build ON them.
The gap between tools and value was bridged by humans. We computationalize this gap.
Better Practice, Not Just Developer
AI as Developer
- Writes code in isolation
- No project awareness
- Forgets everything per session
AI as Project Manager
- Orchestrates scope, context, quality
- Multi-project coordination
- Persistent organizational learning
The hard problem is context construction, not code generation.
Worker Layer: Engineering Discipline
delta-templates define per-project principles, skills, and commands
Rules as Code
Behavioral directives in .claude/rules/ that AI loads automatically every session.
Skills & Commands
Reusable capabilities: /start-task, /task-done, /code-review, /decompose.
File-Based State
Git-tracked proposals, inbox state machine, Linear as source of truth for tasks.
Quality Pipeline
Intent → Demand Review → Plan → Plan Review → Execute → Code Review. Three gates before code ships.
Interactive + Autonomous
Same SOP, different supervision level
Day Mode (Interactive)
Human guides each step. Reviews plans before implementation.
Night Mode (Nightrun)
Quality gates replace human checkpoints. Fleet of parallel workers.
Every workflow improvement benefits both modes. You choose a supervision level, not a tool.
PM Layer: Multi-Project Orchestration
Meta-Repo
Linear + GitHub Integration
Cross-project visibility via unified task tracking and automated PR workflows.
Worktree Isolation
Each task gets its own git worktree. Parallel nightrun workers, zero interference.
Team Project Management
Daily/weekly reports aggregate progress across all projects and portfolio students.
The Autonomy Risk
As AI autonomy increases, so does the blast radius of mistakes
Over-Trust
AI confidently produces plausible but wrong output. Bugs ship to production. The more autonomous the system, the later mistakes are caught.
Under-Trust
Humans review everything "just in case." AI capacity is wasted. The team gains autonomy on paper but not in practice.
How do you know which tasks are safe to automate and which need human oversight?
Without systematic assessment, you are guessing.
SOTIF-Inspired Safety Review
Systematic automation risk assessment, borrowed from automotive safety
ARIL Scoring
Automation Risk Impact Level per module. Quantifies reliability risk of AI-driven changes.
Fractal Tree Traversal
Bottom-up risk propagation through CLAUDE.md hierarchy. Leaf modules aggregate to root.
Capability-Aware Gating
Match task complexity to current AI capability. Route high-risk tasks to human review.
The system evaluates its own automation safety before executing.
The Context Crisis
AI starts every session with amnesia
Zero Memory
Each session starts from scratch. Decisions made yesterday are invisible today. The AI re-discovers what it already learned.
Knowledge in Heads
Team knowledge lives only in human memory. When someone is unavailable, the knowledge is gone. Unscalable.
Inconsistent Workers
Different AI workers get different context depending on who prompted them. Quality varies wildly across sessions.
The hard problem of AI-assisted development is context construction, not code generation.
Code generation is solved. Knowing what to build, why, and how it fits -- that is the gap.
Context Is All You Need
Four sources converge into unified task context
Task
Context
Every AI worker starts with the full picture. No amnesia. No guessing.
Fractal CLAUDE.md
The right context at the right time -- never too much, never too little
T1: Session Start
Root CLAUDE.md auto-loads. Project identity, children map, data flow. Always available.
T2: Module Access
Per-module rules load when touching that module. Conventions, insights, local patterns.
T3: On Demand
Deep docs read explicitly. SPEC.md for architecture rationale. SCOPE.md for strategy.
Learning System
Manual corrections compound into organizational memory
Tier 1: Local Rules
Per-project behavioral directives. Immediate effect next session. Project scope
Tier 2: Project Insights
Cognitive patterns, auto-refined to 1500 chars max. Project scope
Tier 3: Team Principles
Cross-project review criteria. Requires human approval. Team-wide
Bitter Lessons Pipeline
Every code review finding is captured, tagged, and auto-retrieved when the same pattern recurs. Bug fixes trigger root-cause reflection.
This is not RLHF. It is explicit, auditable organizational learning.
Three-Layer Emergence
Quality emerges from disciplined context construction at three time scales
Structured Redundancy Minutes
Multi-agent planning. Iterative review. Gate-fix retry loops. More compute per task = more reliable output.
Evolutionary Selection Hours
Plan + code review select the best approach. Failed attempts inform future decisions.
Compound Learning Weeks
Human feedback routed to 3-tier system. The system improves through the work it enables.
Better prompts help one task.
Better structure helps every task, forever.
Progressive Autonomy
Results & Trajectory
Current
| Metric | Now | Phase B | Phase D |
|---|---|---|---|
| PR Approval | ~50% | 80% | >95% |
| Tasks/Week | 3-5 | 10-15 | 30+ |
| Human Role | Code + Plan | Outcome Review | Hypothesis |
AI4AI4X
Use AI to build AI infra. Use AI infra to transform industries.
The Loop
PM-DELTA manages its own development. Every improvement immediately improves the process that builds it.
Deep Cultivation
Value from domain depth, not breadth. AI+Finance, AI+Education -- each domain compounds learning.
The Flywheel
Compound Growth
The system that ran 100 tasks is meaningfully better than the one that ran 10.
AI Cost-Effectiveness Map
Internal framework guides domain selection: which industries benefit most from AI automation today.
Join Us
Building a team of domain-deep AI practitioners
AI+X PhD Students
Combine AI methods with domain expertise: finance, education, healthcare, manufacturing.
Full-Time Engineers
Build AI infrastructure that compounds. Shape the future of AI-native development.
Research Partners
Cross-institutional collaboration. Real data, real systems, real impact.
Not just model builders -- domain practitioners who wield AI as infrastructure.
PM-DELTA
AI-Native Development Infrastructure
github.com/SII-DELTA