Deterministic AI infrastructure engineered for autonomous agent systems.
I’m Jurie Smit. Over the past two and a half years at Phoenix VC, I engineered an integrated infrastructure stack to coordinate agent teams, route models, and enforce cryptographic task leases across production ventures.
- 1 open source
- Retort (MIT)
- 3 in beta
- Sluice, Baton & Docket
- 3 in alpha
- Cognitive Mesh, Deck & Whet
Invented from Real-World Scars
Celladore wasn't designed in a vacuum or brainstormed on a whiteboard. Every single tool was built in direct response to something breaking in production across real ventures. Filter by domain or toggle between deep-dive anatomy breakdowns, compact summaries, and the origin timeline:
Rogue autonomous storytelling agent loops triggered recursive generation cascades, racking up massive token bills in a single week with zero budget visibility.
Direct client-side LLM calls without rate-limiting proxies, virtual key tenancy, or hard spend guardrails.
Built Sluice and Docket to enforce per-workspace virtual keys, automated fallback routing, and real-time Azure billing threshold alerts.
Running 20 parallel IDE windows across Cursor, Windsurf, Claude Code, and Gemini CLI simultaneously during an intense real-world property renovation caused agents to clash, duplicate work, and clobber each other's edits.
Lack of a centralized task lock manager, cross-repo state reconciliation, and cryptographic execution leases.
Created Baton — the shared task graph with cryptographic execution leases, 3-tier MCP protocol provider, and bidirectional repo YAML sync.
Edge simulation and hardware sensor monitoring suffered from browser tab GC stutter, high memory overhead, and dropped WebSocket connections during live operations.
Web browser UI architectures are fundamentally unsuited for low-latency native IPC and real-time sensor streams.
Engineered Deck — a native Tauri 2.0 shell paired with a high-performance .NET 9 sidecar and type-safe Rust contract bridge.
Unconstrained LLM swarms drifted into subtle hallucination loops and untested architectural regressions without deterministic boundary controls.
Unchecked natural-language agent instructions lacking formal cognitive containment and strict, automated quality gates.
Formulated the Cognitive Mesh architecture, enforcing automated AST audits, typed contracts, and Retort deterministic quality gates.
Ingesting raw multi-megabyte WhatsApp export files, voice notes, and PDF attachments blew prompt token budgets with redundant formatting noise.
Naive raw-text chunking preserves non-semantic document noise, polluting embedding vectors and agent context windows.
Built Mill (formerly xtox) to parse PDFs, DOCX, LaTeX, and codebase directories into token-optimized ASTs with mathematical formula preservation.
Multi-agent PR generation and code-review crews crashed halfway through 2-hour refactoring jobs due to transient API drops, losing uncommitted progress.
Stateless script-based agent loops fail catastrophically when any intermediate HTTP call or process terminates.
Engineered Whet (formerly CodeFlow) with Temporal durable state machines and CrewAI agent crews that resume effortlessly across retries and restarts.
Jurie Smit
Systems Engineer · AI Architect · CTO at Phoenix VC
I’m a systems engineer based in South Africa. Over the past two and a half years at Phoenix VC, I built and iterated Celladore’s modular toolchain to orchestrate autonomous agents, prevent execution deadlocks, and control token spend across production venture applications like NeuralLiquid, Nexamesh, and Mystira.
I author research on Cognitive Sovereignty and designed the Cognitive Mesh architecture. Every tool in this ecosystem was written to solve a real failure mode experienced while coordinating multi-agent pipelines in production.
Talk Directly with the Engineer
No sales tiers or generic account managers. Get direct technical guidance on multi-agent architectures, sovereign routing gateways, and custom LLM adapters.
Eight Composable Infrastructure Tools
Architected into three clear operational domains — from open source runners to managed gateways and desktop shells.
The foundational open-source orchestration engine powering the Celladore ecosystem. Retort coordinates polyglot agent teams, encapsulates modular packaged skills, and enforces strict quality gates before any agent output is merged. Open source and built natively for enterprise Windows environments.
npm install @celladore/retort && retort run --team shipCoordinate specialized agent teams against complex objectives with dynamic delegation, shared memory, and strict role separation.
Distribute reusable agent capabilities, tools, and system prompt harnesses across repositories without duplicating configurations.
The nervous system of multi-agent development. Baton synchronizes tasks across repositories via bidirectional YAML sync, issues cryptographic execution leases (HMAC-SHA256) to eliminate subagent collisions, captures evidence receipts, and tracks historical agent review quality over MCP.
npx @celladore/baton-cli sync --manifest ./tasks.yamlIssue time-bounded HMAC-SHA256 execution leases to autonomous agents, ensuring strict single-owner mutation rights and zero collision risk.
Maintain repository issue manifests in human-readable YAML while syncing real-time task states with the central Baton graph.
Unified Multi-Modal Infrastructure Pipeline
From raw document ingestion to verifiable model routing and live cost telemetry — Celladore provides the full enterprise agent execution pipeline with zero vendor lock-in. Switch between the step-by-step Carousel walkthrough, deep-dive Accordion, or high-level Flow Grid.
Explore the 5 interconnected layers powering deterministic agent cognition and sovereign execution.
Multi-Modal Ingestion & AST Parsing
Lossless conversion of multi-format inputs into token-optimized ASTs
Transforms unstructured PDFs, DOCX, LaTeX, WhatsApp transcripts, and whole GitHub repositories into clean, token-optimized AST representations. Eliminates prompt context bloat before agents ever see the data.
Prompt context window exhaustion and garbled RAG chunking artifacts.
mill.parse(stream, { flattenAst: true, preserveMath: true, targetTokens: 4096 })- Input Format:
- Raw PDF, DOCX, LaTeX, Codebase Tree
- Output Format:
- Token-Optimized AST & Structured Markdown
- Protocol Tier:
- Streaming AST Parser
Container Topology & Representative Logs
Explore the production microservice architecture, representative structured audit events, and cryptographic task lease schemas deployed across Azure Container Apps and Azure AI Foundry.
Production Microservices Architecture
Verifiable runtime infrastructure deployed across Azure Container Apps and Azure AI Foundry to coordinate agent teams without collisions.
retort-agent-worker
Executes multi-turn agent dispatch loops with strict verification gates and non-interactive subagent isolation.
- Runtime Environment
- Azure Container Apps (Linux x64)
- Interface Protocol
- stdio / RPC
- Core Technology
- Rust / Node.js 22
- Availability SLA
- 99.9% Production SLA
Production API Contracts & Typed SDKs
Interactive, copy-ready cURL, TypeScript, and Python snippets backed by typed contracts across all seven ecosystem tools: Retort, Sluice, Baton, Docket, Deck, Whet, and Mill.
Unified Developer Interface
Declarative agent team orchestration with typed quality gates and skill runners.
import { RetortEngine, defineTeam, qualityGate } from "@celladore/retort";
const team = defineTeam({
name: "code-review-crew",
skills: ["ast-analysis", "security-audit", "diff-review"],
gates: [
qualityGate("unit-tests", async (ctx) => ctx.runTests()),
qualityGate("lint-clean", async (ctx) => ctx.runLinter()),
],
});
const engine = new RetortEngine({ config: team });
const execution = await engine.execute({ prNumber: 29 });
console.log(`Execution complete: ${execution.status}`);Telemetry & Routing Benchmarks
Target SLA benchmarks, LiteLLM gateway routing latency metrics, and sovereign enterprise tenant isolation models.
Performance & Privacy Benchmarks
Simulated latency breakdown, target SLA benchmarks, and tenant privacy isolation architecture.
Azure Container Apps multi-region SLA
LiteLLM virtual-key routing layer
No prompt or completion storage
Dedicated Azure Key Vault isolation
Transparent Tiers & Zero Lock-in
Install the Retort orchestration engine locally for free under MIT, or deploy the complete tool stack — including Mill's source-available ingestion engine — to your private cloud.
Retort Engine
Run multi-agent dispatch loops completely offline or with your own LLM API keys.
npm install @celladore/retortCLInpx @celladore/retort initSetupNote: npm packages are pending publication. See GitHub repositories for current installation.
- ✓Retort orchestration runner (Rust / TypeScript)
- ✓Typed JSON schema contracts & local tool execution
- ✓Zero telemetry tracking, runs 100% locally
Celladore Cloud & Gateway Stack
Deploy Baton task lease brokers, Sluice LiteLLM gateways, and Docket budget guardrails into your own Azure or AWS cloud subscription.
- ✓All 7 tools configured for Azure Container Apps
- ✓Cryptographic A2A task lease coordination
- ✓Multi-provider LiteLLM routing & automated fallbacks
- ✓Direct architecture onboarding with Jurie Smit
Direct Answers to Technical Questions
Is this production-ready?
The core stack has been running real multi-agent workloads, comms analysis, and storytelling engines for over two years across Phoenix VC ventures. It is built for real developer workflows, not pitch decks.
What about open source sovereignty?
The foundational orchestration engine (Retort) is MIT-licensed — fork it, run it locally, and inspect every line. The multi-modal parser (Mill, formerly xtox) is source-available under its own terms.
Why Azure & .NET?
Built natively on Azure Container Apps and Azure AI Foundry via Microsoft for Startups Founders Hub. This pairs enterprise Azure scale and billing APIs with .NET 10, Rust, and TypeScript.
Why work with Celladore?
You get high architectural velocity and direct access to the systems engineer who designed and built the stack, eliminating bureaucratic support layers and bloated vendor lock-in.
Architecting Autonomous Agent Teams?
Whether you are integrating agent teams into your development pipeline, exploring sovereign cloud deployments on Azure, or collaborating across venture architectures:
Direct Access to the Founder
Direct communication with Jurie Smit. Reach out for technical deep dives, private beta deployments, or strategic advisory.
jurie@celladoresystems.com