Illicium V4 — Sovereign AI Engineering Platform

Your developers trust Copilot.

Copilot never read your rulebook.

BANK SDK - APPROVED

[email protected] - SAFE

Approved manifest - CICD enforced - Signed

→

DEVELOPER
TRUSTED
THIS

COPILOT SESSION - LIVE

Generating code...

import lodash from 'lodash'

[email protected] PULLED
CVE-2024-1337 CRITICAL
PROTOTYPE POLLUTION
Copilot used public weights - no SDK lookup

DEVELOPER CODES

→

COPILOT SUGGESTS

→

PUBLIC LIB PULLED

→

CVE INJECTED

→

AUDIT FAILS

→

CLIENT COMPLAINS

→

RELEASE BLOCKED

🛑

CVE FLOOD

Copilot pulls vulnerable public libraries. Security blocks the release.

→ Copilot adds [email protected]. CVE found. Release blocked. Client waits 2 weeks.

🧠

ZERO MEMORY

Every AI session starts blank. No knowledge of bank rules or approved SDKs.

→ New dev opens Copilot. Uses public logger instead of bank SDK. Audit catches it 3 months later.

⚠️

COMPLIANCE DRIFT

AI doesn't check if code follows bank architecture standards.

→ Copilot picks Express.js — not approved. Passes unit tests. Fails architecture review. 4 weeks rework.

"WE ALREADY HAVE SAST / CONTAINER SCANNING / RENOVATE — WHY V4?"

REACTIVE — FIND PROBLEMS AFTER

CODE WRITTEN

→

BUILD

→

SAST SCAN

→

🔴 CVE!

→

TICKET

→

FIX

→

RESCAN

SAST — false positive noise

Container — scans after image built

Renovate — bumps version, doesn't patch source

Hundreds of alerts. Weeks of rework. Developers hate it. CVE enters first, then you chase it.

PREVENTIVE — BLOCK PROBLEMS BEFORE

CODE WRITTEN

→

SDK GATE

→

✅ APPROVED ONLY

→

BUILD

→

DEPLOY ✓

SDK Gate — only approved libs enter

FOSS Vault — patched at source code

Zero noise — nothing bad gets in

Zero CVE alerts. Zero rework. Developers love it. The gate prevents, not the scanner that chases.

SAST and Renovate are firefighters. V4 is the fireproof building. You still run scanners — but they find nothing.

AI is replacing software engineers.

55%

of production code is already
AI-generated

— GitHub Copilot Report

2–3yr

until AI handles
most engineering tasks

— Anthropic, OpenAI, DeepMind

0

banks with a governed
AI engineering pipeline

— the gap Illicium fills

👁️

NOBODY CAN READ CODE

AI writes everything. Developers lose the skill. When AI is wrong — who fixes it?

📦

BINARY IS UNAUDITABLE

AI generates → compiles to binary → regulators ask what's running. No trail.

👤

HUMAN REVIEW AT SCALE

Banks need humans accountable. AI generates thousands of lines. Who reviews without governance?

🎯

AI CAN BE HACKED

Prompt injection, adversarial inputs, poisoned data — AI code is an attack surface.

The question is not whether AI replaces engineers. It is who governs the AI that replaces them.

Where does your engineering
intelligence go?

PATH A — BUY SaaS

✗ Your code leaves the bank perimeter

✗ Vendor learns your patterns & workflows

✗ Licensing fees — every year, forever

✗ AI burns tokens scanning vendor CVEs repeatedly

✗ Blind spots — code outside your wall

💰💰💰

They own your intelligence. You pay forever.

PATH B — BUILD SOVEREIGN

✓ Code stays inside the bank perimeter

✓ Your architect writes the libraries — bank owns IP

✓ Build once — own forever, zero licensing

✓ AI reuses your kernel — fraction of tokens

✓ Zero blind spots — everything inside your wall

💰→✓

You own the intelligence. It compounds.

Every SaaS vendor you use gets smarter about your bank.
Every project on Illicium makes your bank smarter about itself.

One sovereign loop. Everything governed.

THE SOVEREIGN LOOP — HOW EVERY AI OUTPUT IS GOVERNED

AI OUTPUT

INTERCEPT

VALIDATE SDK

ENFORCE OPA

SIGN CRYPTO

EMIT ✓ APPROVED

Every AI output. Every team. Intercepted, validated, enforced, signed, emitted. No exceptions. No bypasses.

Not a concept. Running today.

TODAY — V3 (JENKINS) · FINANCIAL ENGINEERING DEVOPS

Kelvin

1 ARCHITECT

→

V3 FRAMEWORK

shared governance

→

✓ Team A — quantitative models & libraries

✓ Team B — end-user computing

✓ Team C — platform & runtime services

✓ Team D — AI model training & serving

Chain DAG pipelines · Monorepo subprojects · Zero manual gates

ISTIO ✓

→

HARBOR ✓

→

KAFKA

→

CLICKHOUSE

→

+ MORE

V4 — NEW FRAMEWORK (REPLACES JENKINS · REWRITE FROM SCRATCH) · BANK-WIDE

V4 gets bank approval → every DevOps team can adopt

WHAT EVERY TEAM GETS OUT OF THE BOX:

🔧 Default Templates

Ready-made pipeline patterns

🖱️ Drag & Drop Canvas

Visual workflow builder

🧪 Debug Sandbox

Input/output test stages

⚡ Own Tool Choice

nx, dagger.io, Jenkins, custom

Teams design their own workflows — like customising Jira. Default templates get them started fast. Framework governance stays consistent underneath.

IT DevOps · Risk DevOps · Finance DevOps → all inherit, all design their own way

✗ Jenkins NOT APPROVED by bank CICD

→

V4 removes Jenkins. Teams keep their tools. Framework is bank-approved.

Approve once. Win every platform.

WITHOUT FRAMEWORK — WHAT MOST BANKS DO

Every new product = build from scratch

NEW PRODUCT

→

NEW TEAM

→

NEW CODE

→

NEW APPROVAL

→

💰💰💰

× repeat for EVERY product. AI rewrites from zero each time = more tokens, more cost, more risk.

WITH ILLICIUM — APPROVE ONCE, REUSE FOREVER

Every new product = plug into the approved kernel

NEW PRODUCT

→

SAME KERNEL

→

SAME POLICY

→

NO RE-APPROVAL

→

✓ LIVE

AI reuses the kernel patterns = fraction of tokens, fraction of cost, zero new procurement.

Libraries & Framework

Built by Kelvin's Special Force Team — Rust, WASM, DAG. FOSS-inspired, CVE-patched. Bank owns the IP.

BUILT ONCE

Platform Products

V4 SDLC, Canvas, AI Ops, FOSS Vault — each platform inherits the same approved policy & compliance rules.

APPROVED ONCE

Team Rules

Each team gets default templates — then customises their own workflows, like Jira. Team A, B, C, D — all reuse, all different.

TEAM DESIGNS

"WHY NOT JUST USE PREFECT / TEMPORAL / KARPENTER?"

FOSS ENGINES = WORKERS

Prefect, Temporal → execute workflows

Karpenter → scales compute

n8n → connects nodes visually

They run what you give them. They don't ask if it's allowed.

vs

ILLICIUM = GOVERNANCE ABOVE ALL

→ Decides what is authorised to run

→ Enforces bank policy before execution

→ Signs every artifact cryptographically

→ Remembers every decision (sovereign memory)

Then tells Prefect / Karpenter / n8n what to execute.

We don't replace execution engines. We govern them. They become workers inside the framework.

ONE APPROVAL → THE FACTORY THAT BUILDS EVERYTHING ELSE

V4 SDLC

THE FRAMEWORK

→

BUILDS:

FOSS Vault

first delivery ✓

Canvas

drag-and-drop DAG

AI Ops

when LLM approved

ML Platform

dev env governance

Without V4, there is no FOSS Vault, no Canvas, no AI Ops. V4 is the factory. One approval → every product.

How it actually works. V3 today.

Same pattern powers V4 SDLC, Canvas, FOSS Vault, AI Ops. One control plane. One orchestrator. Engines execute — Illicium decides.

Canvas + AI. The future is already here.

Business users build governed workflows with AI — no code, no risk, any cloud.

ILLICIUM CANVAS — quarterly_risk_report.flow

🔧 DEV MODE 🔒 RUNTIME ▶ Run in Sandbox 🧪 Debug

✓ SDK ENV: DEV TARGET: AWS

NODES drag →

📥 Source

⚙️ Transform

📊 Output

🔀 Branch

🔧 Custom

TARGET ▾

AWS EKS

GCP GKE

Cloud Build

Ali Cloud K8s

Model Exec

Runtime Svc

On-Prem

WORKFLOW CANVAS — drag nodes from sidebar or click to debug ↓

📥 Get Risk Data

risk_db · DEV · APAC

TARGET: EKS

→

⚙️ Filter & Aggregate

by region · quarterly

TARGET: GKE

→

📊 Load Analytics

analytics_dw · append

TARGET: On-Prem

➕ Drop a node here to extend workflow

🧪 DEBUG — click a node above ↑ IDLE

📥 INPUT

-- click a node --

📤 OUTPUT

-- click a node --

✓ SDK 🔒 PROD/DEV ✍️ SIGNED 🛡️ AI SANDBOX ONLY each node targets different cloud

🤖 AI ASSISTANT

USER

"Get customer risk data for APAC, aggregate by quarter, load into analytics."

AI

Built 3-node workflow.
Worker: AWS K8s (sandbox).
✓ SDK validated · PROD safe

USER

"Looks good. Push to PROD."

⚠️ GOVERNANCE

PROD requires approval.
Sent to Kelvin. Waiting...

SAME CANVAS — DIFFERENT PRODUCTS — DIFFERENT TEAM ACCESS

n8n CANNOT DO THIS ✗

PRODUCT: Risk Analytics

✓ risk_db · ✓ market_data

✗ customer_pii · ✗ PROD write

AI: LangChain → risk scoring → report

Target: EKS + On-Prem

PRODUCT: AI Research

✓ model_registry · ✓ LLM-as-Judge

✗ production_db · ✗ customer data

AI: train → eval → guardrail → staging

Target: GKE sandbox

PRODUCT: Finance Reporting

✓ finance_dw (read-only) · ✓ export

✗ write access · ✗ external APIs

AI: query → validate → report → sign

Target: On-Prem only

Design behaviours. Ship immutably.

Each subproject gets its own behaviour. Debug any stage. Test before you ship.

ILLICIUM SDLC — harbor-registry-chain.template

🔧 DEV MODE 🔒 RUNTIME ▶ Run 🧪 Debug

✓ SDK │ SIT → DEV → UAT →🔒→ PROD │ TARGET: EKS

BEHAVIOURS drag →

🟦 Node.js

🟦 Golang

🟨 Python

🟪 Java

🟩 FOSS Clone

🟥 Docker

🔧 Custom

STAGES drag →

📥 Clone

🔨 Build

🧪 Test

📦 Package

🚀 Publish

➕ Custom

🔒 L1 ONLY
Clients use RUNTIME.
Team rules restrict access.

TEMPLATE — MONOREPO SUBPROJECTS

🟦 harbor-portal/ Node.js v2.4.1 ✓ PASS

npm install → npm build → npm test → docker build

🟩 harbor-core/ Golang FOSS Clone v1.8.3 ✗ BLOCKED

git clone (FOSS) → go build → go test ✗ → docker build

🧪 Debug: click to inspect test output — identify root cause before deploying

🟪 db-exporter/ Golang v1.2.0 ✓ PASS

go build → go test → docker build

➕ Drop a behaviour here to add subproject, or a stage to extend pipeline…

🏷️ VERSIONING

Semver auto-tagged per subproject. Each release gets its own version.

📂 AUTO-DETECT

Select git repo → framework scans project files → applies behaviour. Override with .illicium.yaml

🧪 DEBUG

harbor-core blocked at go test — click to inspect and fix before deploy.

DAG CHAIN ORDER

L1 harbor-portal · base → L2 harbor-core · db-exporter

🤖 AI ASSISTANT

SOVEREIGN

ALERT: harbor-core go test failed — TestRegistryAuth panic: nil pointer

AI: Investigating SIT environment READ-ONLY
▸ Pulled SIT logs for harbor-core
▸ Root cause: mock registry client not initialised — missing NewTestClient() at line 47

AI: Fix ready + config update:
+ client := NewTestClient()
~ .illicium.yaml timeout: 30→60s

🔒 ROLE PERMISSION CHECK

✓ Kelvin (L1) — delegate AI to push
✓ AI can hotfix SIT/DEV configs
✗ UAT/PROD requires approval

AI: Permission granted. Pushed fix to SIT.
✓ PR #189 created · Tests pass · SIT deployed

HARNESS HAS AIDA™ — SO WHAT?

Their AI runs on their cloud. Your code leaves.
V4 AI is sovereign — learns your patterns, governs by your roles.

Ask AI to debug, configure, or fix...

Send

🔧 DEV MODE

Edit behaviours · Debug bugs · Deploy to SIT/DEV/UAT

🔒

🚀 RUNTIME

Locked · Immutable · Crypto signed · Every release identical

V4 ships a default pipeline.

Every project gets CI/CD on day one. Teams customise. AI fixes what breaks.

ILLICIUM PIPELINE — harbor-core.pipeline.yaml | ● LIVE RUN #247

DEFAULT STAGES

✓ Git Clone

✓ Setup Env

✓ Version

✓ Build

✓ Package

✓ Test

✓ Publish

◉ Scans

○ CD

CUSTOMISE

+ Add Stage

+ Add Target

+ Override

📥 Clone 2s

→

⚙️ Setup 5s

→

🏷️ v1.8.4 1s

→

🔨 Build 18s

→

📦 Package 8s

→

🧪 Test 12s

→

🚀 Publish 4s

→

🚢 CD ✓

✅ CI/CD COMPLETE — 50s

Artifact published. Deployed to SIT for testing.

Devs test in SIT/DEV playground while scans run — no waiting.

◉ ASYNC SCANS RUNNING

◉ SAST ◉ Container ✓ SonarQube ✓ FOSS

Non-blocking — AI auto-fixes when results arrive

🔒 RELEASE GATE

No bypass.
Scans must pass before
release or hotfix to
UAT / PROD

🤖 AI AGENTIC AUTO-FIX — Agent Swarm · Loop Engineering

Spawns agent pods via Karpenter — each pod runs in isolated env, targets any cloud. Subprocesses for dependent tasks. Parallel execution — not sequential.

1. SCAN DETECTS

CVE-2026-3891

golang.org/x/net

→

2. SPAWN AGENT

▸ Karpenter → new pod
▸ Isolated sandbox env
▸ Target: EKS

→

3. AGENT FIXES

                - x/net v0.17.0

                + x/net v0.23.0

                  go mod tidy ✓

→

4. VERIFY (subprocess)

✓ Spawns CI/CD subprocess
✓ Full pipeline on fix branch
✓ Tests + scans pass
✓ Pod terminates

→

5. PR CREATED

PR #248

CI/CD proof attached
✓ Verified by pipeline
✓ Ready for review

DEPLOY TARGETS:

EKS GKE On-Prem Cloud Build Ali Cloud K8s Model Exec Runtime Svc

🤖 AI ASSISTANT

SAST: Critical CVE detected in harbor-core go.mod — golang.org/x/net v0.17.0

AI: CVE-2026-3891 affects HTTP/2 handler. Upgrade to v0.23.0 patches it. No breaking API changes. Writing fix...

AI: Updated go.mod + go.sum. Ran go mod tidy. All tests pass. Created PR #248.

Container: All clear ✓ No vulnerabilities in base image.

V3 WOULD HAVE...

⏱ Blocked CI for 15+ min
📋 Created 47 Jira tickets
👨‍💻 Developer reads each one
🔧 Manual fix, hours to days

V4 DELIVERED

⚡ CI never blocked
🤖 AI read the CVE advisory
🔧 Auto-wrote the fix
✓ PR ready in 90 seconds

📋 DEFAULT

9 stages out of the box. CI/CD on day one.

⚡ FAST

Deploy SIT/DEV instantly. Scans never block devs.

🔒 SAFE

No release or hotfix without scans passing. Zero bypass.

What this unlocks.

⚡

New team joins → productive in days

They inherit the framework. AI already knows the patterns. No 3-month onboarding. No per-team rewrite.

🧑‍💼

Business users build workflows

Canvas + AI assistant. Non-technical users create governed pipelines. No developers needed for routine work.

📋

Regulator calls → answer in minutes

ClickHouse sovereign memory has every decision, every approval, every signature. No forensics needed.

🔄

AI gets smarter about YOUR bank

Every project governed = patterns learned. The flywheel accelerates. After 12 months, no competitor can replicate this.

🛡️

AI can never wipe PROD

PROD/DEV enforced at the framework level. AI works in sandbox only. Human approves promotion. No accidental disasters.

📦

New FOSS tool → governed in days

AI reads the FOSS code, generates an Illicium version, SDK validates, CVE-patched. Available to every team. Not months of procurement.

AI is moving fast. The bank that governs it first doesn't just survive — it accelerates.

Beyond ReAct. The sovereign agentic mesh.

ReAct is already standard. V4 is built for what comes after.

2023 — STANDARD

ReAct

Reason + Act loop

Single agent. No memory.

→

2024 — EMERGING

Multi-Agent

CrewAI · AutoGen

Agents collaborate. No governance.

→

2025 — NOW

Cognitive Arch

Memory · Plan · Reflect

Claude Code, Devin. Still SaaS.

→

NEXT — V4

Sovereign Mesh

Governed · Memory · Crypto

Bank owns it. Nobody else has this.

SOVEREIGN AGENTIC MESH — HOW V4 GOVERNS AUTONOMOUS AGENTS

🧠 PLAN

🤖 DELEGATE

⚡ EXECUTE

🔒 VALIDATE

💾 REMEMBER

🔄 EVOLVE

Agents own tools

Agents negotiate

Persistent memory

Self-improve

Crypto signed

BUILD FROM SCRATCH — LangChain + OpenAI SDK

NEW AI APP

→

NEW APPROVAL

→

NEW SECURITY

→

6+ MONTHS

V4 INTERNAL SDK — PRE-APPROVED FOUNDATION

NEW AI APP

→

SAME SDK

→

INHERITED

→

DAYS ✓

🔄

Release Mgmt

AI agents review, test, approve, deploy. Autonomous release pipeline.

🎧

Help Desk

Agent triages, routes, resolves from sovereign memory. Governed responses.

📋

Kanban AI

Agents move cards, assign reviewers, enforce SLA. Teams set their rules.

🖱️

Canvas IDE

Business users drag-and-drop. Agents build the governed pipeline underneath.

Your team presented ReAct. V4 is already built for what comes after. The governance layer doesn't care which pattern the agents use — it governs them all.

V4 vs the market. Same vision. Sovereign.

Harness ($3.7B) is the world's leading SDLC DevOps platform — CI/CD, developer portals, AI testing, security, cost management. 15 products. Thousands of companies. This is V4's direct competitor.

n8n is the leading open-source workflow automation — drag-and-drop DAG canvas. This is Canvas's direct competitor.

Both are SaaS or FOSS with no bank governance. Illicium delivers both — sovereign, air-gapped — plus 3 capabilities neither will ever build ↓

CAPABILITY	MARKET	BANK HAS	ILLICIUM
CI/CD	Harness	IT CICD	Sovereign pipeline
Workflow	Temporal	Model Execution (IT)	Governs what Model Execution runs
Runtime	Karpenter	Runtime-as-a-Service (bank)	Intent layer above Runtime-as-a-Service
Dev Env	Codespaces	Machine Languages	Session SDK enforcement
Portal	Backstage	--	Integrate or sovereign rewrite
Canvas	n8n	--	Drag-and-drop DAG sovereign
AI Govern [STAR]	NONE	NONE	LLM-as-Judge - LangChain eval - OPA
Memory [STAR]	NONE	NONE	ClickHouse sovereign - never leaves
FOSS Library Vault [STAR]	NONE	V3 (Istio, Harbor live)	Istio, Harbor, Kafka, ClickHouse + more — CVE-patched, bank-owned assets
Data	Harness cloud	Partial	100% your perimeter

[STAR] = No market equivalent. No bank equivalent. Unique to Illicium.

Your architect is building a sovereign FOSS vault — Kafka, ClickHouse, and more — CVE-patched at source, brought into the bank as institutional assets. No vendor. No licensing. No supply chain risk.

Harness is $3.7B. It cannot be used by a bank. Illicium delivers the same vision -- sovereign, air-gapped, with the AI governance and FOSS vault that banks actually need.

12 months from now.

When this moves forward.

400+

PROJECTS GOVERNED

One kernel. Every team governed. No manual gates.

Model Exec + Runtime

ABSORBED

Governed workers. Illicium is the intent layer above both.

CANVAS + ML

ENVIRONMENT LIVE

Visual drag-and-drop DAG live. Machine Languages sessions controlled.

AI OPS

ACTIVE

LLMs bank-approved. Sovereign loop running 24/7.

FOSS VAULT

EXPANDING

Istio, Harbor, Kafka, ClickHouse — CVE-patched. More every quarter.

BANK-WIDE

ADOPTION

IT DevOps, Risk, Finance — all teams on one framework. One standard.

One architect. Every AI workflow. Governed.

The knowledge accumulated in 12 months cannot be purchased from any vendor. Ever.

One more thing.

V3 migration codes itself.

📄

V3 JENKINS
PIPELINE

→

🤖

AI READS
V3 PATTERN

→

⚡

GENERATES
V4 DAG WORKER

→

✅

SDK VALIDATES
DEPLOYED

Every V3 Jenkins pipeline migrates to V4 automatically. The Loop Engineering SDK reads the existing pattern, generates the equivalent V4 DAG worker, validates it, and deploys. No migration team. No big-bang project. Pipeline by pipeline.

Jenkins becomes optional. The migration writes itself.

The libraries are being written now.

The framework architecture is designed.

V3 proves the pattern works in production.

4 teams governed. FOSS assets flowing. Chain pipelines live.

The only question left is ownership.

The standard is being built.
The question is who owns it.

WAR ROOM BRIEF

TACTICAL DETAIL -- AVAILABLE ON REQUEST

INTERNAL POLITICAL LANDSCAPE

Model Execution is owned by IT. Replacing it directly creates conflict. Illicium governs above it -- not against it.

Runtime-as-a-Service is a bank infrastructure product. Illicium positions as the intent layer, not a competitor.

Machine Languages team controls developer environments. SDK enforcement is a capability add, not a constraint.

Jenkins rejection is structural -- V4 resolves it by removing the dependency entirely.

3-PHASE STRATEGY

Phase 1 -- GET APPROVED: Secure funding. Ship L1 kernel. Pass bank CICD review. Prove V4 replaces Jenkins cleanly.

Phase 2 -- PROVE CAPABILITY: Onboard two additional teams. Demonstrate Model Execution governance layer. Ship Canvas MVP. Activate AI Govern.

Phase 3 -- AI CODES MIGRATION: Loop Engineering SDK reads Model Execution patterns. AI generates Illicium DAG workers. Migration becomes self-executing. Model Execution becomes legacy on its own timeline.

COMPLIANCE WALL DETAIL

Jenkins not on bank approved CICD list. IT cannot use it in standard pipelines. V3 requires Jenkins -- this is the scaling ceiling.

V4 L1 Kernel uses Rust/WASM, no Jenkins dependency. Full sovereign pipeline from day one.

OPA Rego policies approved once at L2 and inherited -- no per-team compliance review per product.

BUDGET ARGUMENT BREAKDOWN

One architect. One kernel. Marginal cost per additional team is near-zero -- pattern inheritance, not per-team build.

Harness enterprise license equivalent: $500K+ per year. Illicium: one-time build cost, perpetual IP ownership.

Compliance re-approval per new SaaS product: significant overhead. Illicium: zero -- Bank Armor inherited automatically.

Institutional AI knowledge in ClickHouse sovereign memory compounds. Cannot be purchased later -- must be grown from day one.