# Néo Schobert

> Applied AI engineer: knowledge graphs, LLM agents and industrial software. R&D at Infineon in Munich, founder of INOVA System, graduating from IMT Atlantique in December 2026. Available from January 2027, in France or elsewhere in Europe.

Néo Schobert is a French computer science engineer (IMT Atlantique, graduating December 2026, with a bachelor's in pure mathematics). He works on GraphRAG agents, ontologies and knowledge graphs, and on industrial software that replaces PLCs with connected, remotely updated systems.

## Current roles
- Infineon Technologies AG: R&D Intern (final-year internship) (Jul – Dec 2026). Designed and shipped Ontology Explorer, a six-service platform to import, explore, align and query enterprise ontologies.
- INOVA System: CEO & Founder (Alongside his studies). An industrial OS to replace PLCs: distributed software embedded on machines, connected and maintained remotely through regular updates.
- SalsaDeRue: Dance instructor & association director (2023 – present). Running a Latin dance association of 150+ members, with full financial and operational responsibility.

## Selected projects
- Argus — real exploitability of C++ CVEs: An agent that answers the real question: can this vulnerability actually be reached in this code? It builds the call graph of a C++/CMake project with tree-sitter, fetches CVEs and public exploits, traces data flow back from every risky call, then a local model writes an exploitability report. The code never leaves the server.
- AgentForge — a team of agents that ships pull requests: Open an issue, get a pull request back: a LangGraph agent graph (supervisor, architect, coder, tester, reviewer) implements, tests and reviews the task, for up to three iterations. Each task runs in its own Kata Containers microVM, and API keys are never exposed to the agents.
- INOVA System's server: a company that runs itself: A server he hosts himself, with all of the company's data kept on-site, and most of its operations automated: an internal GraphRAG over our documents, bookkeeping, e-mail, setting up new projects, all the way to generating the software for an automated machine with built-in simulators.

## Contact
- Email: neo@neo-schobert.com
- LinkedIn: https://www.linkedin.com/in/neo-schobert
- GitHub: https://github.com/neo-schobert
- Website: https://www.neo-schobert.com
- CV (PDF): https://www.neo-schobert.com/CV_Neo_Schobert.pdf

## More
- [Full profile in plain text](https://www.neo-schobert.com/llms-full.txt): experience, Infineon work, INOVA System, projects, skills, education, languages.


---

## Experience

### Infineon Technologies AG — R&D Intern (final-year internship)
Jul – Dec 2026 · Munich, Germany · Corporate Supply Chain Innovation

Designed and shipped Ontology Explorer, a six-service platform to import, explore, align and query enterprise ontologies.
- GraphRAG agent: 14 native tools (hybrid search, SPARQL, graph traversals, news watch), four effort levels up to parallel sub-agents, cited and self-verified answers: 93% correct on a 42-question benchmark.
- Python alignment engine: 12 matchers, LLM verifier, SKOS/SSSOM exports; first of 5 engines compared, F1 0.936 on OAEI Anatomy.
- Versioned news watch and bitemporal fact graph, exposed to other teams via a read-only API.
Stack: RDF/OWL, SPARQL, GraphRAG, LLMs, TypeScript, React, Python, PostgreSQL, Docker

### INOVA System — CEO & Founder
Alongside his studies · France

An industrial OS to replace PLCs: distributed software embedded on machines, connected and maintained remotely through regular updates.
- Runtime on Linux panel PCs, open controllers and IIoT sensors instead of proprietary PLCs.
- Fleet cloud: OTA updates, supervision, remote access and regulatory traceability.
- End-to-end ownership: client needs, design, development, deployment, billing and sales strategy.
Stack: Industrie 4.0, Edge, OTA, EtherNet/IP, OPC-UA, MQTT, Linux, Cloud

### MA Solutions — R&D Engineering Intern
Jun – Sep 2025 · France · machines for hospital laundries

Full IIoT architecture to replace legacy PLCs: remote maintenance, OTA updates and machine control.
- Linux Panel PC + EtherNet/IP I/O board replacing PLCs: 10+ confirmed orders.
- Web platform (client portal + admin dashboard) for fleet management, remote access and regulatory traceability.
- Retrofit of existing machines with an OPC-UA server on a VPS, with no hardware change.
Stack: EtherNet/IP, OPC-UA, Linux, Node.js, React, SQL, VPS

### SalsaDeRue — Dance instructor & association director
2023 – present · Finistère, France

Running a Latin dance association of 150+ members, with full financial and operational responsibility.
- Mobile app published on Google Play and the App Store, downloaded by 200+ people: memberships, ticketing, events and notifications.
- Weekly salsa classes, workshops and social nights.
Stack: React, React Native, TypeScript, Node.js, Firebase

### Serralharia Rolo — Production intern
Jun – Jul 2024 · Porto, Portugal

Hands-on metalwork internship: cutting, welding, assembly. Full immersion in Portuguese within a team of craftsmen.

## Infineon Technologies — Ontology Explorer (his final-year R&D internship)

At Infineon, knowledge is spread across ontologies from many domains and a mass of documents. The reference supply chain ontology, the Digital Reference, has more than 1,200 classes: far too many for anyone to keep in their head. Every team invents its own vocabulary, alignment is manual, and general-purpose LLMs answer confidently while citing nothing.

Nothing is asserted that can't be traced: every answer points back to an IRI or to a sentence that actually exists in a source.

### GraphRAG engine

Not a plain RAG: an agent runs the investigation with a toolbox over the ontology, writes from the evidence alone, then checks its own answer before sending it.
1. Understand: A follow-up becomes a standalone question, concept names are resolved in the open folder, and the agent asks for clarification when the question is ambiguous. Tools: ask_user.
2. Plan: In deep and research modes, a plan is written before acting and shown to the user.
3. Collect: One agent drives the whole toolbox, calls several tools at once and relaunches the search when needed. In research mode, an orchestrator splits the question across four sub-agents in parallel. Tools: search_concepts, concept_card, schema_property, sparql_query, graph_shortest_path, graph_longest_chain, graph_top_degree, alignment_correspondences, evidence_read.
4. Write: From the evidence alone: every fact cited by its IRI, in the language of the question, with a diagram of the subgraph used and its Turtle export.
5. Verify: Numbers and IRIs cross-checked against the evidence, re-read by a judge model in advanced modes. Every diagram edge is looked up in the ontology: kept, flipped or dropped. An invented identifier triggers a targeted rewrite.

Effort levels:
- Fast: The original pipeline: condense, hybrid retrieval (65% vector, 35% lexical), routing to lookup, SPARQL or graph traversal, then writing. The cheapest, the most predictable.
- Standard: The agent: native tools called in parallel, relaunchable search, step-by-step sourced answer, mechanically verified.
- Deep: The agent with an explicit plan, twice the budget, a judge re-read and a rewrite when needed.
- Research: An orchestrator splits the question, four sub-agents dig into it in parallel, then synthesis and verification. Several times the cost: an explicit choice, never a default.

Benchmark (Measured on 42 Digital Reference questions):
- Original pipeline: 60% correct, 31% hallucinations, 13.6 s median, $0.026 per question
- Standard agent: 93% correct, 7% hallucinations, 29.6 s median, $0.053 per question

### Architecture
- frontend: The web UI: 3D/2D graph, workspace, chat and news watch. It only ever talks to explorer.
- other teams: The engines are stateless and the watch is exposed read-only: other applications reuse them without touching the database.
- explorer: The UI's API and the only service that writes: accounts, folders, ontologies, SPARQL, news watch. It checks permissions, then mints a ticket for every engine call.
- alignment: Alignment engine: 12 matchers and an LLM verifier, SKOS and SSSOM exports. It keeps nothing once the answer is sent.
- chatbot: The GraphRAG agent: knowledge arrives with the question and is forgotten once the answer is written. No cache, no stored conversation.
- news: The news watch for other teams: GET routes only, a read-only SQL account limited to an allowlist of tables.
- PostgreSQL: Accounts, folders, ontologies, documents, news watch and vectors. Engines get no table privileges: only five read-only functions, opened by a ticket.

### Measured results
- 93 %: correct answers on a 42-question benchmark (60% with the original pipeline)
- 31 → 7 %: hallucinations, from the original pipeline to the agent
- 0.936: alignment F1 on OAEI Anatomy (best published: 0.941)
- #1 / 5: alignment engines compared, on every test set
- 3,000+: automated tests across the six services

### Research written during the internship
- State of the art: GraphRAG over ontologies: Why plain vector RAG falls short, why “LLM → SPARQL” alone is a trap, retrieval techniques, matching, ontology embeddings.
- Shapes of graphs: 18 graph data models compared on the same fact: RDF, property graphs, hypergraphs, n-ary, temporal…
- Temporal knowledge graphs: Bitemporality, LLM event extraction, temporal KG forecasting applied to supply chain risk.

Makeathon: Internal 2026 Makeathon: a multi-agent orchestrator with one expert agent per ontology, parallel sub-tasks (SPARQL, APIs, search), answers with provenance and confidence, and ontology edits proposed by agents then approved by a human. He owned the architecture and technical coordination.

## INOVA System, the company he founded and runs

The PLC has barely changed in forty years: frozen logic, on-site interventions, no usable data. INOVA System replaces it with an industrial OS: distributed software embedded on every machine, connected, and maintained remotely through regular updates.

- Layer L4 · Applications: The client portal and fleet dashboard: every machine, its status and version, in real time. (Client portal, Fleet dashboard, Traceability)
- Layer L3 · Fleet cloud: The single control point: supervision, remote access and rolling out updates to the whole fleet. (OTA rollout, Supervision, Remote access)
- Layer L2 · Edge runtime: On every machine, INOVA OS replaces the PLC: it drives the I/O and runs business logic locally, even offline. (Linux panel PC, Business logic, Offline-ready)
- Layer L1 · Equipment: The machine stays the same: the OS plugs into its EtherNet/IP and Modbus I/O and its sensors, with no hardware swap. (EtherNet/IP I/O, Modbus, IIoT sensors)
- One update, the whole fleet: Published once in the cloud, the new version flows down to every machine, installs and checks itself, with no technician on site.

Born in the field: the IIoT architecture designed in R&D at MA Solutions, already chosen for 10+ orders, turned into a product.
Website: https://industrie.inova-system.com

## SalsaDeRue, the association he runs, and its mobile app

SalsaDeRue is the Latin dance association he runs in Brittany: 150+ members and classes in several cities. Its app gives them their schedule, class videos, messaging and membership card.

- Schedule: Weekly classes, live cancellations
- Videos: Moves from class, by level
- Messaging: Groups, teachers and carpools
- Member card: QR code, also in Apple Wallet
- Parties & workshops: Agenda, sign-up and ticketing
- Shop: Delivered right to class
200+ downloads. Designed, built and shipped solo, from the back-end to both stores.
Google Play: https://play.google.com/store/apps/details?id=fr.salsaderue.app
App Store: https://apps.apple.com/fr/app/salsaderue/id6806458439

## Projects

### Argus — real exploitability of C++ CVEs
IMT Atlantique · Final-year project (Procom) · team of 4 · 2025–26

An agent that answers the real question: can this vulnerability actually be reached in this code? It builds the call graph of a C++/CMake project with tree-sitter, fetches CVEs and public exploits, traces data flow back from every risky call, then a local model writes an exploitability report. The code never leaves the server.
Result: LLM-judge benchmark with cross-review: up to 8.8/10 after consensus, and a correct 0 for the CVE-free project.
Stack: Python, FastAPI, tree-sitter, Ollama, SQLite, Docker, Next.js

### AgentForge — a team of agents that ships pull requests
Personal project · open source (MIT) · 2026

Open an issue, get a pull request back: a LangGraph agent graph (supervisor, architect, coder, tester, reviewer) implements, tests and reviews the task, for up to three iterations. Each task runs in its own Kata Containers microVM, and API keys are never exposed to the agents.
Result: Fully self-hosted: Forgejo Git, LangFuse tracing, a React dashboard and a streaming “project lead” chat.
Stack: Python, LangGraph, LangFuse, Kata Containers, Docker, FastAPI, React

### Website and app for a 1,500+ member club
Client project · confidential · 2025–26

A professional website and its app, installed straight from the browser as a home-screen shortcut: online ticketing, built-in member messaging and event management, for a 1,500+ member club that runs two events a month.
Stack: Next.js, React, Firebase, PWA

### Prompt Share — a shared AI chat in VS Code
Personal project · VS Code extension, open source · 2026

Several people in one chat with the host's Copilot agent, from VS Code or a browser; the host approves every change and the agent stays confined to the project.
Stack: TypeScript, VS Code API, WebSocket

### Multi-agent job search assistant
Personal project · self-hosted · 2026

Twelve Claude Code sub-agents: job monitoring, scoring, application documents and a browser copilot, over a 33-table database.
Stack: TypeScript, Claude Code, Agents, PostgreSQL, Docker

### An adaptive, LLM-powered language tutor
Personal project · 2026

For learning any language: exercises generated on your weak spots and corrected by an LLM, a CEFR level adjusted continuously, spaced repetition, long-term RAG memory and voice conversation.
Stack: TypeScript, Python, Mistral, RAG

### INOVA System's server: a company that runs itself
INOVA System · self-hosted · 2026

A server he hosts himself, with all of the company's data kept on-site, and most of its operations automated: an internal GraphRAG over our documents, bookkeeping, e-mail, setting up new projects, all the way to generating the software for an automated machine with built-in simulators.
Result: Automatic backups, single secure sign-on, and the entire configuration versioned in one repository.
Stack: Self-hosting, GraphRAG, LLMs, Docker, ZFS, Authentik, Tailscale

### Graph neural networks for linear programming
IMT Atlantique · Research project, ML track · solo · 2024–25

Representing a linear program as a constraint–variable graph and training a message-passing network to predict feasibility, boundedness and the optimal solution, following recent literature.
Result: 0.8 accuracy on feasibility generalisation.
Stack: Python, PyTorch, GNN

### Ticket buyer prediction
IMT Atlantique · Big Data · solo · 2025–26

A PySpark pipeline on real data from his association's platform (800 users, 30 events): 22 features, Random Forest vs GBT, then a ranked list of the members most likely to buy, exported to target outreach.
Result: AUC 0.987, with an honest look at possible target leakage.
Stack: PySpark, MLlib, Firestore, Docker

## Skills

### AI & LLMs
- GraphRAG (used at: Infineon · INOVA)
- LLM agents & tools (used at: Infineon · INOVA · AgentForge)
- LangGraph (used at: AgentForge)
- Embeddings & vector search (used at: Infineon · INOVA)
- Local & self-hosted models (used at: INOVA · Argus)
- LLM automation of business workflows (used at: INOVA)
- LLM-as-judge evaluation (used at: Infineon · Argus)
- PyTorch (used at: IMT Atlantique)

### Knowledge graphs
- RDF · OWL (used at: Infineon · INOVA)
- SPARQL (used at: Infineon · INOVA)
- Graphs built from documents (used at: Infineon · INOVA)
- SHACL (used at: Infineon · INOVA)
- SKOS · SSSOM (used at: Infineon)
- Ontology alignment (used at: Infineon)
- Oxigraph (used at: Infineon)

### Software
- TypeScript (used at: Infineon · INOVA)
- Python (used at: Infineon · Argus)
- React · Next.js (used at: Infineon · INOVA)
- React Native · Expo (used at: SalsaDeRue)
- Node.js · FastAPI (used at: Infineon · Argus)
- PostgreSQL (used at: Infineon · INOVA)
- Firebase (used at: SalsaDeRue)

### Infra & industry
- Docker (used at: Infineon · INOVA)
- Linux · self-hosting (used at: INOVA)
- OTA updates (used at: INOVA · MA Solutions)
- EtherNet/IP · Modbus (used at: MA Solutions)
- OPC-UA · MQTT (used at: MA Solutions)
- TLS · SSO (used at: Infineon · INOVA)

## Education

### IMT Atlantique (2023 – 2026)
Master's in Engineering, Computer Science, machine learning track · graduating December 2026
A broad engineering degree, with a machine learning specialisation and projects built around LLMs: agents, model-as-judge evaluation, deep learning and knowledge graphs.

### Université de Bretagne Occidentale (2023 – 2024)
Bachelor's in Fundamental Mathematics, alongside first year of engineering school
Linear algebra, Lebesgue integration, commutative rings, group theory, arithmetic.

### CPGE Montaigne (2020 – 2023)
MPSI then MP preparatory classes · admitted to the ENS
CPGE is France's selective preparatory track for the top schools; MPSI then MP is its maths-and-physics stream. Three years of mathematics, physics and computer science at a demanding pace, with weekly oral exams and tests, ending in the national competitive exams, through which he was admitted to the ENS (École normale supérieure) and to IMT Atlantique.

## Languages

- French: Native
- English: Fluent — IELTS 7.5 (C1)
- Spanish: Intermediate (B1)
- Portuguese: Intermediate (B1)