SHEET 02 / 06WORK · PLATFORMS & FIELD ARCHIVE

Three platforms, one thesis: faster code only moves the bottleneck.

AI copilots made writing code cheap, so the constraint moved downstream — review, handoffs, compliance. But most enterprise AI stays trapped in silos. Each platform below replaces a siloed step with one governed agentic layer — and each is sized like a P&L, not a demo.

~€520M

Annual value at stake · base case — full model on Sheet 06 →

~7:1

Blended value-to-cost ratio across the portfolio

<9 mo

Payback, including the adoption ramp

40%+

Faster engineering cycles

METRIC FLAGS: E — EVIDENCED I — ILLUSTRATIVE / DIRECTIONAL TARGET — every number below carries one.

open dossier --id=aura

AURA — the whole SDLC as one platform. I

DOSSIER 01 / SHEET 02

Agentic Unified Reactive Agile platform. A custom agentic framework with an abstraction layer over agents, agile-role agents inside VSCode / IntelliJ, and MCP servers wiring Jira, ServiceNow, Confluence, MBSE, Rhapsody and LINK. Built around MCP and A2A, with multi-LLM routing and distributed inference over vLLM + Ray. Brownfield-first.

PLAN

BUILD

AURAgoverned layer

REVIEW

SHIP

OPERATE

⟲ OPERATE FEEDS PLAN — ONE GOVERNED LOOP BENEATH EVERY STAGE

FIG. 01 — AURA GOVERNS THE CYCLE, NOT ONE SILO

Problem

AI copilots made writing code cheap.
The constraint moved downstream — review, integration, tech debt, security.
Each SDLC stage is still optimised in its own silo.

Solution

An enterprise agentic SDLC platform with an abstraction layer over agents (incl. BMAD).
Agile-role agents inside VSCode / IntelliJ plugins.
MCP servers wire Jira, ServiceNow, Confluence, MBSE, Rhapsody, LINK.

Unique Value

Platform thinking for the whole SDLC — not point fixes.
Velocity, quality and maintenance improve together, under one governed layer.

Customers & Channels

Software & systems engineering teams; agile squads, architects, V&V.
Delivered via IDE plugins + internal platform onboarding.

Key Metrics

Backlog velocity · review cycle time.
Escaped-defect & tech-debt rate.
Autonomous-observability coverage.

Unfair Advantage

Deep OEM toolchain integration (MBSE / Rhapsody / LINK).
Built-in IAM, GuardRails and compliance — an agentic SDLC the enterprise can approve.

2.1×BACKLOG THROUGHPUT

−45%REVIEW CYCLE TIME

−38%ESCAPED DEFECTS

>90%PIPELINE OBSERVABILITY

THE AUTONOMOUS PRODUCTION LINE Seven role-based agents run in parallel where they can, in handoff where they must: requirement analysis → coding ⇄ data + unit testing → review → merge → E2E validation with self-healing. The requirement-stage Blueprint persists immutably through to validation, so every downstream agent verifies against original intent. Playwright E2E failures trigger recursive loop-backs into coding, data or unit-test phases. Zero-touch from backlog ticket to shipped feature.

open dossier --id=customer-portal

Customer Portal — agents that hand off like a real team. I

DOSSIER 02 / SHEET 02

An agentic orchestration ecosystem: every SDLC phase is an agent, and the loop runs from requirements to end-to-end without a human carrying context across boundaries. The healing agent patches loop-backs without a full rollback.

ANALYSTreqs · semantic match

CODINGimplements

DEPLOYmerge · triage

HEALINGpatches loop-backs

⟲ HEALING → ANALYST · PATCH LOOP-BACK, NO FULL ROLLBACK

FIG. 02 — FOUR AGENTS, ONE CLOSED LOOP

Problem

SDLC phases stay disconnected.
Work is handed off manually between requirements, code, deployment and fixes.
Context, time and traceability are lost at every boundary.

Solution

A chain of specialised agents, one per phase.
Analyst — requirements + semantic match vs the existing codebase → Coding → Deployment → Healing.
Healing patches loop-backs without a full rollback.

Unique Value

The SDLC automated through intelligent agent handoffs.
A closed loop from requirements to E2E — not disconnected copilots.

Customers & Channels

Enterprise engineering orgs running a multi-phase SDLC.
Platform / DevEx teams; delivered via an internal portal.

Key Metrics

Requirement-to-E2E lead time.
Handoff context loss · auto-heal success rate.
Rollback avoidance.

Unfair Advantage

Self-healing without a full rollback.
Semantic codebase matching at intake.
Roadmap to multi-repo, cross-project scale.

−55%REQUIREMENT-TO-E2E LEAD TIME

~70%AUTO-HEALED, NO ROLLBACK

4SDLC PHASES ORCHESTRATED

3×REPOS UNDER ONE LOOP · ROADMAP

open dossier --id=genui

GenUI — review and create in one AI tool. I

DOSSIER 03 / SHEET 02

Vehicle-OS UX guidelines are intricate enough to overload designers, and compliance takes multiple rounds of manual review. GenUI collapses guideline review and prototyping into one dual-mode tool: Reviewer Mode files compliance tickets automatically; Generator Mode turns text or sketches into standards-aligned HTML/CSS prototypes, integrated with Figma and Adobe XD.

DESIGN INPUTtext · sketch · figma

REVIEWERguideline check

TICKETSauto-filed

SHIP

DESIGN INPUTtext · sketch · figma

GENERATORsketch / text → code

PROTOTYPEHTML / CSS

SHIP

ONE INPUT DRIVES BOTH MODES — REVIEW CATCHES IT, GENERATION FIXES IT

FIG. 03 — DUAL-MODE: AUTOMATED REVIEW + DESIGN-TO-CODE

Problem

Vehicle-OS UX/UI guidelines are intricate — cognitive overload for designers.
Compliance needs multiple rounds of manual review.
No automation in the design-to-ship loop.

Solution

A dual-mode SaaS tool.
Reviewer Mode — automated guideline-compliance review that auto-creates tickets.
Generator Mode — text or sketch to HTML/CSS prototypes; Figma and Adobe XD plugins.

Unique Value

Review and create in one AI tool, built for the vehicle OS.
Compliance checking and prototyping in a single workflow.

Customers & Channels

Vehicle-OS UX/UI designers, design-system owners, product teams.
Delivered as SaaS + Figma / Adobe XD plugins.

Key Metrics

Adoption & retention.
Task-completion time · review iterations.
NPS · CSAT.

Unfair Advantage

The only tool tailored to the OEM's design guidelines.
Dual reviewer + generator, AI-powered end to end.

30%YEAR-1 ADOPTION TARGET

−50%TASK-COMPLETION TIME

−40%REVIEW ITERATIONS

+50NPS · CSAT 85%+

PATTERN · 01

Diagnose the real constraint

Every win started by naming the true bottleneck, not the obvious one. AI made code cheap — so review, handoffs and compliance became the constraint.

PATTERN · 02

Build a platform, not a point fix

Point fixes only move the bottleneck; platforms remove it. AURA, the Customer Portal and GenUI each replace siloed steps with one governed agentic layer.

PATTERN · 03

Prove it with P&L economics

Each initiative is sized bottom-up — cost pool × evidenced uplift, net of a verification tax — so the number is defensible, not aspirational.

git clone --depth=full

Same thesis, shipping in the open.

SEC 05 / SHEET 02

Three repositories where the autonomous-SDLC, declarative-agent and on-device-AI patterns live in public — Rust and C++, deployable from a laptop to a cluster.

zenith RUST · MIT

Nine specialised agents — Architect, Developer, Tester, Reviewer, DevOps, Research, plus a 3-stage Playwright pipeline that plans, generates and self-heals browser tests. Three-provider failover across Gemini, Claude and Azure OpenAI; every action gated through an approval dashboard; 207 tests passing. Runs entirely on your own hardware.

YOUR AI ENGINEERING TEAMgithub.com/plushpluto/zenith ↗

agentfile RUST · MIT

A declarative format for portable AI agents — one file defines tools, memory, multi-model orchestration, pipelines and the service layer. Ten LLM providers. Run locally, serve as HTTP, or deploy to Kubernetes via the official Helm chart and AgentDeployment CRD. SPAWN · PARALLEL · AGGREGATE for multi-agent topologies.

DOCKERFILE, FOR AGENTSgithub.com/plushpluto/agentfile ↗

KLLM C++ · C · CMAKE

Language models injected at the kernel level — pseudo-level layering, on-device small LMs, a finite-state machine for state and recovery, and Sentinel AI for real-time monitoring, jailbreak prevention and self-healing. The edge-AI thesis from the flight-computer days, applied to the operating system itself. Targets Android, Ubuntu Touch, web-OS.

KERNEL-LEVEL LANGUAGE MODELSgithub.com/plushpluto/kllm ↗

mount /archive --era=2013..2025

Field archive.

SEC 06 / SHEET 02

The hardware built, the demos flown, the dashboards shipped — NTU, SwarmX, Republic Polytechnic, Vishwa Dynamics, and two 2025 refreshes. Stills first, then the tapes.