Work
Systems work and the research behind it.
Case studies, open-source work, and active research tracks — the what and why, not the proprietary how.
Case studies
Systems work, framed by architecture judgment.
Problems, constraints, principles, outcomes, and lessons learned.
01
Platform architecture
User Platform Architecture
A SaaS platform core designed around clear client, BFF, domain, identity, notification, and integration boundaries.
Architecture Lens
system.mapProblem
Many SaaS products need a durable user, authentication, authorization, notification, and integration foundation before product teams can move quickly with confidence.
Difficulty
The challenge was balancing a small-team operating model with architecture choices that could still support clear boundaries, future protocol evolution, and reliable validation.
Principles
Keep business rules in the domain layer, use the BFF to shape client experience, document architectural decisions, and design integration paths that can evolve from simple protocols to stronger internal contracts.
Outcome
A platform foundation with documented architecture, operational runbooks, manual validation flows, and an evolution path for service communication.
Lesson
Platform architecture is not just service decomposition. It is the discipline of making boundaries, operations, and future change legible to both humans and AI-assisted engineering workflows.
02
Design system architecture
Design Engine Kit
A theme-driven design system architecture for reusable SaaS interfaces, generated layouts, and defensible enterprise UI composition.
Architecture Lens
system.mapProblem
Product teams often duplicate UI decisions across dashboards, workspaces, generated pages, and customer-specific experiences, making brand adaptation and layout reliability expensive over time.
Difficulty
The hard part was separating brand data from UI mechanisms while keeping components resilient across density, theme, layout, content length, and responsive edge cases.
Principles
Separate mechanism from brand data, bind components to semantic design contracts, make layout resilience part of the system, and treat composite components as product infrastructure rather than isolated widgets.
Outcome
An architecture for a theme engine, semantic token layer, layout system, defensive UI rules, and enterprise-ready composite component patterns.
Lesson
A serious design system is not only a component library. It is a runtime contract between brand, layout, accessibility, data density, and product evolution.
03
AI-assisted engineering workflow
Spec Generator
A planning system that translates plain-language product intent into structured engineering specifications, phased action plans, and reviewable delivery constraints.
Architecture Lens
system.mapProblem
Product ideas often move into implementation before intent, technical constraints, acceptance criteria, and engineering governance are made explicit.
Difficulty
The challenge was making AI-generated specifications useful to both non-technical stakeholders and engineers without leaking sensitive internal process details.
Principles
Preserve user intent, translate requirements into engineering concepts, make constraints visible, separate planning from execution, and keep humans responsible for reviewing AI-produced artifacts.
Outcome
A product design for generating editable spec artifacts, technical translation tables, phased implementation plans, exportable documentation, and traceable rules-version context.
Lesson
AI-assisted engineering becomes more valuable when the system produces structured intent, constraints, and acceptance criteria before asking an agent to implement.
Open source
Public work you can read and run.
Selected public repositories behind the governance and platform work. Early projects are marked as such.
Orvena
v0.9.0
williamlabdev/Orvena
An OS-enforced task-scope governance runtime for AI coding agents: bounded execution, verify gates, and frozen evidence.
ABL-GATE
Technical report
williamlabdev/aine-abl-gate
A pre-registered, paired ablation of enforced versus advisory completion gates on LLM coding agents, published with sanitized evidence.
AINE Registry
v1.6.0
williamlabdev/aine-registry
A read-only CLI that discovers projects, artifacts, dependencies, impact, preflight risk, and agent context across multi-root portfolios.
AINE Control Plane
v0.2.0
williamlabdev/aine-control-plane
Provider-neutral, read-only contracts and reference primitives for policy evaluation, access control, and approvals over portfolio evidence.
SaaS Forge
Public snapshot
williamlabdev/saas-forge
A Go SaaS backend published to be read and run: modular monolith, domain REST API with a GraphQL BFF, transactional outbox, and pluggable authorization.
Skill Security Linter
Early
williamlabdev/skill-security-linter
A neutral trust gate for agent skills: turns a skill into a permission inventory and risk score so CI can block high-risk installs.
AI Gap Radar
Early
williamlabdev/ai-gap-radar
Research notes on fast-rising AI repositories: why they spread, which gaps they leave open, and which spin-offs are worth building.
De-AI Tone
Early
williamlabdev/de-ai-tone
Editorial rules for Chinese and English marketing copy that replace generic AI phrasing with verifiable specifics. A quality check, not a detector-evasion tool.
Research tracks
Open questions behind the architecture work.
Durable problems across AI-native engineering, governance, context, runtime systems, and adjacent decision systems.
2025 - Now
Active
Runtime Systems
ArchitectureInvestigating how AI-assisted engineering environments should preserve continuity, make autonomy observable, and support responsible recovery from uncertainty.
Research Question
What must exist around AI agents before they can participate safely in long-running engineering work?
Current Finding
The essential design challenge is continuity and accountability, not raw autonomy.
2024 - Now
Active
Context Engineering
ResearchExploring how engineering context remains useful, trustworthy, and durable across long-running work while protecting sensitive implementation choices.
Research Question
How should engineering context remain useful across sessions, systems, and decisions without exposing sensitive mechanisms?
Current Finding
Context quality depends on intentional structure, refresh discipline, and human-legible boundaries.
2026 - Now
Active
Engineering Governance
EnterpriseDesigning governance models for AI-assisted engineering where decisions, constraints, and approvals remain visible and enforceable.
Research Question
How can teams increase AI-assisted autonomy while preserving review, risk management, and ownership?
Current Finding
Governance works best when it is part of the engineering system, not a separate approval layer bolted on afterward.