Data → AI Agents → Physical AI
A public lab for intelligent systems.
Research, learning, experiments and builds, documented in public. Data foundation since 2014, direction agents and physical AI.
Projects
AI Decision Copilot
A tabular analytics agent — profiling, investigation, explainable prediction, scenarios — not a chat-over-CSV demo.
AI Automation Agent
A single triage agent — closed-set classification, extraction, score, human review. No irreversible external actions.
AI Knowledge Copilot
A grounded document assistant — retrieval before generation, visible excerpts, confidence signals. Not a generic chatbot.
Web3 Analytics Dashboard
An on-chain dashboard from CSV exports (Dune or otherwise) — KPIs, series, cohorts. Blockchain as a data terrain, not the Lab’s identity.
Latest experiments
A measurable vision baseline from a data seat
UCI digits 8×8, frozen split, test accuracy: majority 10.03% vs centroid 92.48%. Not a VLM.
First Physical AI notes from a data seat
Entry lab: map VLM / VLA / ROS2 / simulation. The first pixel run is the 8×8 digits lab.
An automation agent with no irreversible side effects
v2, same 20 cases: classification 100%, false autonomy 15% (down from 60% in v1).
Can an LLM profile a dataset without inventing stats?
Versioned A/B/C run: 8 rows, tool, extracted claims. B error = 0. Frozen trace: C = A.
Currently exploring
MCP
A tool protocol for agents. Exploring: no public MCP server in this Lab.
ROS 2
Robot middleware. Exploring: no public ROS 2 workspace in this Lab.
Simulation
Synthetic world before the real one. Exploring: no published robot or physics sim.
VLA
Vision-language-action. Exploring: reading, no deployed robot policy.
Signals
All topics · 2026-10-18
Each column has its own unit. This table does not rank topics. Editorial status stays manual.
| Topic | GitHub | arXiv | HN | HF |
|---|---|---|---|---|
| AI Agents | 7689 | 5740 | 4039 | 1406 |
| AI Evaluation | 7793 | 1178 | 1465 | 1558 |
| Physical AI | 622 | 284 | 125 | 105 |
| VLA | 543 | 2338 | 28 | 3352 |
| AI Coding | 7336 | 1481 | 886 | 41419 |
AI Agents
Builder and research activity is accelerating faster than evaluation evidence.
AI Evaluation
Without a failure metric, an AI system is only a story. The topic is rising; public suites are still scarce.
Physical AI
VLM, VLA, simulation and world models. Uneven maturity, still-unstable vocabulary.
VLA
Vision-language-action demos are multiplying faster than reproducible stacks.
AI Coding
Coding assistants are a daily tool. The Lab uses them; it has not published one.
Writing
Agent vs workflow
Three public systems, three autonomy levels. Not everything needs to be an agent.
What I am learning in intelligent systems
A public journal: data → agents → physical AI. Evidence on one side, still-empty nodes on the other.
Physical AI from a data perspective
What changes when you move from tabular pipelines to perception, action and simulation — without pretending to be a robotics engineer.
Trajectory
Guillaume Verbiguié, Paris. Senior Data Analyst at AdvanThink since 2023, after Ipsos (2018–2022). Data practice since 2014 — public sector, banking, retail, research. Public labs and projects show what comes next: agents, RAG, Physical AI as a direction.