Solo-built agentic AI ecosystem from Switzerland on a 100W NVIDIA GB10 Blackwell desktop supercomputer. Three independent agentic projects: cognitive robotics (Unitree Go2 + Isaac Sim 5.1 + GR00T N1.7), local-first BI (DuckDB + Claude NL→SQL), and LLM-reasoning EDR cybersecurity. Production-grade, no cloud lock-in.
Author: Minh-Tam Dang — Infinity Cloud Sàrl, Geneva, Switzerland · infinitycloud.ch
Solo builder. Twenty years of program management at scale (Pernod Ricard, ICRC). Stepped back from corporate roles in 2023 to pursue causes that matter — education, robotics, and AI. A privilege I take seriously. Founded Infinity Cloud Sàrl in Switzerland in 2024.
Hardware footprint: a personal NVIDIA DGX Spark (Grace-Blackwell GB10, 128 GB unified memory, 100 W TDP) under the desk, paired with a Mac M3 Ultra. No cloud rental, no rented GPUs. Every workload runs on hardware I own.
Method: an "agentic farm" of specialized AI agents (STRAT/DEV pairs) coordinated through tmux sessions and structured protocols. Each project gets its own agent team. Sprints are short, deliverables are real, validation is QA-driven. The IEEE-format paper PrivExpensIA: Observational Study of a Multi-Agent Orchestration Framework (September 2025) documents the approach.
The orchestration backbone is MonoCLI — a Python CLI built specifically because no tool existed to give AI agents a persistent, evolving brain. SQLite knowledge base, YAML Playbooks for structured missions, Jedi Rank certification system, Scribe for continuous learning across sessions. 99.8% test pass rate. The same brain that drives the robotics platform also coordinates the project teams.
1. Robotics — GR00T N1.7 + Hy3D pipeline on a 100W desktop supercomputer
Quadruped robotics infrastructure (Unitree Go2 + NVIDIA Isaac Sim 5.1 + ROS2 Jazzy + RL PPO locomotion) running on the GB10 Blackwell. Three-layer architecture: Brain (MonoCLI + LLM planning) → Interface (RobotAdapter, sim/real agnostic) → World (Isaac Sim + OmniGraph). Sim-to-real transfer validated on the physical Go2 (videos in robotics/media/). Sprint 18 ran in fully autonomous mode and proved a complete genetic loop: −18.4% time, −11.8% distance, −17.6% cycles after one round of knowledge distillation. Custom X1 telemetry dashboard for real-time pilot view. Hy3D → Isaac Sim asset pipeline. GR00T N1.7 inference validated on Unitree G1.
Read the LinkedIn article → · See the dashboard, the farm, the robot → · Driven by MonoCLI
2. Agentic ICBI — Orchestrating multiple Claude agents for production-grade BI
Local-first business intelligence: CSV ingestion → DuckDB profiling → LLM-generated semantic layer (YAML) → natural-language-to-SQL with type-aware constraints → adaptive visualization (bar / line / scatter / heatmap / scorecard / orthographic globe). The SYNAPSE/ADN/cipher pattern compresses agent memory across sessions. 32 deliverables in 2 days via full agentic execution from a structured PRD with pre-defined workflows, three architectural pivots absorbed. 440K rows × 94 columns profiled in 3.7 seconds on local hardware. No cloud, no vendor lock-in.
Read the LinkedIn article → · Watch the demo (28 s) → · Same agentic stack as Robotics
3. Cyber Sentinel — LLM-first EDR: detect unknown threats by reasoning, not pattern matching
Defensive cybersecurity for solo operators and small teams. A reasoning-based endpoint detection and response engine that understands what it observes instead of matching signatures. Hardened WireGuard VPN module with strict sudoers scoping, Sentinel monitoring component, methodology for solo-operator threat modelling. Built because existing tooling assumes either a SOC team or a consumer-grade attitude — neither fits the solo professional operating from home with valuable IP on disk.
Read the LinkedIn article → · Built on the same MonoCLI reasoning loop as Robotics and ICBI
Three reasons:
- Public timestamping. Every commit is a dated proof of work. The agentic farm pattern documented in the September 2025 IEEE paper predated mainstream multi-agent frameworks (Agent Teams, etc.) by several months. See
docs/MOULINSART_VS_CLAUDE_CODE.mdfor the chronological comparison. - Reusable patterns. The architecture decisions — local-first, type-aware NL→SQL, three-layer robotics, sudoers-scoped VPN, LLM-reasoning EDR — are reusable. The articles document the why behind each.
- Recruitment-quality artefacts. Engineers and experts who land here can read working architecture, validated benchmarks, and honest postmortems — not marketing.
vision/
├── README.md # This file
├── LICENSE # MIT
├── robotics/
│ ├── README.md # Project overview
│ ├── docs/
│ │ ├── linkedin-article.md # Public-facing article
│ │ ├── architecture.md # 3-layer cognitive model
│ │ ├── hardware.md # GB10 + Mac M3 setup
│ │ ├── methodology.md # Sprint workflow
│ │ ├── results.md # Locomotion + cognitive evolution metrics
│ │ └── roadmap.md # Next phases
│ ├── research/
│ │ ├── vlm_benchmark.md # Vision-language model comparison
│ │ └── compatibility.md # Stack compatibility notes
│ └── media/ # X1 dashboard, agentic farm, Go2 real-condition videos, Isaac Sim
├── agentic-icbi/
│ ├── README.md
│ ├── docs/
│ │ ├── linkedin-article.md
│ │ ├── patterns.md # Technical patterns discovered
│ │ ├── synapse-adn-cipher.md # Compressed memory system
│ │ └── timeline.md # 6-day chronology
│ └── media/ # Screenshots + demo video (333 KB)
└── cyber-sentinel/
├── README.md
├── METHODOLOGY.md # Solo-operator threat model
├── SENTINEL.md # Reasoning-based monitoring
├── VPN_MODULE.md # WireGuard hardening
├── links.md
└── docs/
└── linkedin-article.md
MIT. See LICENSE.
Minh-Tam Dang — Infinity Cloud Sàrl Geneva, Switzerland infinitycloud.ch
Hardware & infrastructure: NVIDIA GB10 Blackwell · DGX Spark · 100 W desktop supercomputer · ARM aarch64 · 128 GB unified memory · Mac M3 Ultra · local-first · edge AI
Robotics: Unitree Go2 · Unitree G1 · Isaac Sim 5.1 · Isaac Lab 2.3 · ROS2 Jazzy · RL PPO · GR00T N1.7 · Hy3D · GEAR-SONIC · sim-to-real · cognitive robotics · assistive robotics
AI orchestration: Claude API · Anthropic · multi-agent · agentic farm · MonoCLI · YAML Playbooks · Jedi Rank · tmux orchestration · STRAT/DEV pairs · IEEE Sept 2025
Data & BI: DuckDB · NL→SQL · semantic layer · type-aware queries · Vue 3 · Observable Plot · SYNAPSE/ADN/cipher
Cybersecurity: LLM-first EDR · reasoning-based detection · WireGuard · sudoers hardening · solo-operator threat model · Sentinel
Origin: Switzerland · Geneva · Infinity Cloud Sàrl · indie developer · applied AI