OCR + LLM extraction
Reads MTC and tensile-test PDFs whatever the mill's template, with a confidence score on every field.
Edge AI · Computer vision · MLOps · Athens, Greece
AI that leaves the lab and runs where your data is born.
Plaixus takes machine learning from notebook to factory floor, farm, robot and satellite. We design the models, build the pipelines and keep them running at the edge, on hardware you already own.
Free 45-minute call. Then a two-week data audit with a written go/no-go before you commit to anything. An engineer, not a sales rep, replies within one business day.
End-to-endfrom data ingestion to model monitoring
On your hardwarean intelligence layer over existing OEM equipment
EU-funded R&DHorizon Europe and cascade-funded delivery in energy and agriculture
In productionTraceLM issues EN 10204 certificates daily at a listed steel producer
Who we are
A deep-tech team that ships models, not slide decks
Plaixus is a Greek deep-tech SME founded in 2022 by an engineering team with research roots. We specialise in industrial-grade machine learning: computer vision and edge AI, MLOps, explainable models, and language models applied to industrial documents. On our platforms or yours.
Most AI projects stall between a promising prototype and a system operators can trust. That gap is our whole business. We work across the full lifecycle: data ingestion and curation, model training and optimisation, edge deployment, monitoring and explainability.
The name comes from plexus: a network of pathways that carries signals between the centre and the extremities. That is what we build: code and data moving gracefully along the cloud-edge continuum.
Book a discovery call-
Production firstEvery model is built to be deployed, versioned, monitored and retrained, not just evaluated.
-
Runs on what you haveCameras, PLCs, SCADA, ROS robots, nanosatellites: we integrate rather than replace.
-
Privacy and compliance by designOn-premise pipelines and federated learning when data cannot travel.
-
Fast, honest feasibilityA two-week data audit tells you whether the problem is solvable before you commit.
Products
TraceLM
Steel certificates, automated
Every steel product sold in the EU will soon need a Digital Product Passport. Most producers still track certificates on paper. TraceLM closes that gap.
TraceLM reads raw mill test certificates (MTCs) and tensile-test reports, extracts the data with OCR and a language model, validates it against built-in rules and issues EN 10204 3.1 and 2.2 certificates automatically. Every match carries a confidence score, and operators correct any OCR mistake before a certificate is generated.
The Digital Product Passport (DPP) plugin makes each product DPP-ready today, so steel processors get ahead of the EU's ESPR rules instead of reacting to them.
It runs in production at an Athens Stock Exchange-listed steel company, on real certificates and real data.
- Ingest MTCsUpload mill test certificates and tensile-test PDFs from any mill, in any layout.
- Extract & validateOCR + LLM extraction, validation rules and a confidence score on every field.
- Generate certificatesEN 10204 3.1 and 2.2 certificates issued per order, traceable to heat and batch.
- Export the passportDPP plugin publishes the product record for ESPR compliance.
Validation & confidence scoring
Built-in rules check every value. Each match carries a confidence score, so review effort goes where it matters.
Full traceability
Every product traced from mill to customer: heat, batch, certificate and order in one record.
DPP-ready today
The Digital Product Passport plugin prepares steel processors for the EU's ESPR rules ahead of the deadline.
Request a TraceLM demo
How the certificate pipeline works
Compliance should be automatic, not manual.
What we build
Six capabilities, one delivery team
We combine them as the problem demands. Most engagements start with a vision or analytics question and end with a monitored pipeline running at the edge.
Computer vision & edge AI
Detection, tracking, counting, anomaly and quality inspection models that run on-device with low latency and no dependency on a stable uplink. Our livestock lameness-detection work under the AGRARIAN programme is a typical example: cameras in the barn, inference at the edge, alerts to the farmer.
- Object detection & segmentation
- Multi-camera tracking
- Visual anomaly detection
- Quantisation, distillation & pruning
- NVIDIA Jetson & industrial gateways
- Offline-first inference
LLM document & telemetry intelligence
Language models applied to documents and machine data, not chatbots: structured extraction from certificates and test reports, retrieval over your own technical documentation, and translation of analyst questions into detection queries over security telemetry. Every output carries a confidence score and goes through human review. Runs on-premise on open-weight models when the data cannot leave. TraceLM is built this way.
- OCR + LLM extraction
- Retrieval-augmented generation
- Tool-using LLM workflows
- On-premise / open-weight LLMs
- Evaluation & guardrails
- Confidence scoring & human review
MLOps & AI pipelines
Reproducible training, automated deployment and continuous monitoring so a model stays accurate long after launch. Drift alerts, retraining triggers and explainability reports your operators can read.
- Kubeflow
- Argo Workflows
- MLflow
- Model registries
- Drift & performance monitoring
Federated & privacy-preserving learning
Train shared models across hospitals, homes or plants without moving raw data. Full compliance with data-protection rules, with the accuracy of a pooled dataset.
- TensorFlow Federated
- PySyft
- Differential privacy
- Secure aggregation
Cloud-native & IIoT integration
We connect models to the systems that matter: DCS, PLC and SCADA on the plant floor, ROS 2 on robots, ESA's NanoSat MO Framework in orbit, with Kubernetes orchestration end to end.
- KubeEdge
- EdgeX Foundry
- Fledge
- ROS 2
- OPC UA / Modbus
Data analysis & decision support
From sensor time series to dashboards and forecasts: predictive maintenance, yield estimation, grid optimisation and explainable models (SHAP, LRP, LIME) your operators can audit.
- Time-series forecasting
- Predictive maintenance
- NLP & document intelligence
- Explainable AI (XAI)
Book a discovery call
Not sure which capability applies? Describe the data you have and we will tell you.
How it fits together
One pipeline across the cloud-edge continuum
Training, optimisation and orchestration happen in the cloud or your data centre. Inference, buffering and control happen at the edge, next to the camera, the controller or the payload computer. Plaixus is the layer in between that keeps the two in sync: models flow down, telemetry and drift signals flow back up.
Because the layer sits on open frameworks, you are never locked into our tooling, and your team can operate it after handover.
Book a discovery callPlatforms
Reference platforms we deploy and adapt
Each one packages a proven pipeline for a specific environment. We use them to start your project weeks ahead, then tailor the last mile to your equipment and data.
-
AIstronaut
Space · edge AI in orbitDeploy and monitor edge AI applications on satellites, combining the power of the cloud with the timeliness of on-board inference for Earth observation and payload autonomy.
ESA NanoSat MO Framework, KubeEdge, EdgeX, Kubeflow -
RoboSense
Robotics & edge fleetsKubeflow and K3s adapted for the realities of the edge, so the same platform runs in a data centre or on the device. Two ways in: build and train models on the platform, or bring a model you trained elsewhere and ship it to any edge device you connect to it. Inference on the device is optimised through NVIDIA Triton.
Kubeflow, K3s, NVIDIA Triton, ROS 2, Argo -
Pledge
Manufacturing & process industryBridge industrial IoT with the cloud and your existing DCS, PLC and SCADA. Removes the data silos that block AI applications in plants and factories.
KubeEdge, EdgeX, Fledge, OPC UA -
Helianthus
Energy & smart buildingsTrain federated models on smart-home and grid data streams. Extends to distribution system operators for substation and RTU data, optimised grid use and predictive maintenance.
Home Assistant, openHAB, HomeEdge -
FederatAId
HealthcareUse multiple healthcare data repositories to train federated AI models in full compliance with data-privacy regulations, without centralising patient data.
Kubeflow, KubeEdge, TensorFlow Federated, PySyft
Sectors
Where our models are running
Physical environments and regulated processes with real constraints: dust, distance, latency, compliance. That is where applied AI pays for itself.
Manufacturing & process industry
Visual quality inspection, predictive maintenance on rotating equipment, and IIoT data unification across legacy control systems.
Steel & metals
Mill test certificates read automatically, EN 10204 3.1 and 2.2 certificates issued with full traceability, and Digital Product Passports ready for ESPR (TraceLM).
Agriculture & livestock
Camera-based animal health monitoring, yield and growth estimation, and offline-capable field devices for farms with weak connectivity (livestock monitoring).
Robotics & autonomous systems
Perception models integrated with ROS 2, deployed and updated over the air to fleets of mobile or industrial robots.
Energy & utilities
Federated analytics on grid and building data, predictive maintenance for substations, and forecasting for optimised grid use.
Space & Earth observation
On-board inference for nanosatellites, so useful insight comes down instead of raw imagery that saturates the downlink.
Cyber security
Language models that translate analyst questions into detection queries over SIEM and IDS telemetry, mapped to MITRE ATT&CK, with the generated query shown for review.
Healthcare
Federated model training across institutions, keeping patient data where it is while improving shared diagnostic models.
Book a discovery call
Your sector is not listed? If the data comes from cameras, PLCs, sensors or certificates, it probably fits. Ask.
Research & funded projects
Research-grade ideas, field-grade delivery
We participate in European R&D as a technical partner that turns concepts into deployable systems. If you are assembling a consortium and need edge AI, MLOps or integration expertise, we would like to hear from you.
Propose a collaborationResearch interests
- Federated & gossip learning
- Delta-sync edge computing
- Quantisation, distillation, pruning
- Conditional computation: MoEs, token selection, early exit
- Vision-language models (VLMs)
- Audio-visual fusion
- Spiking neural networks
- Agentic frameworks: reflection, tool use, planning, multi-agent
- XAI: LRP, SHAP, LIME
- Causal & neuro-symbolic AI
- Neuromorphic computing & SNNs
ELDER: Edge-AI for Lameness Detection in Ruminants
Computer vision that scores individual sheep for lameness from video in the barn. Each animal is tracked across a sequence of frames and the sequence produces one probability, rather than a verdict per frame, with inference running on-device. What it measured in the field.
Edge intelligence on existing OEM equipment
Applying our intelligence-layer approach to industrial and agricultural machinery within the O-CEI Horizon programme.
Edge AI for photovoltaic predictive maintenance
An open-source edge AI platform that detects soiling and loosened bolts on photovoltaic panels, combining image and vibration models running on-device at the array. Selected as one of 15 funded projects from 80 applications across 26 member states in the I-NERGY second open call. What it taught us about intermittent links.
Neuromorphic audio-visual interaction for robotics
Audio-visual speech recognition running as spiking neural networks on a neuromorphic accelerator, driving a robot arm from a single-board computer. One of the first demonstrations of AVSR on this class of processor. Measured energy and accuracy.
Member of the European AI-on-Demand community
Listed on the AI4Europe platform as a provider of data analysis, ML models and ML operations.
How we work
From first call to a monitored system in four steps
Fixed scope at each stage, so you decide with evidence rather than optimism.
Discovery call
We hear the operational problem, look at sample data and agree what success would look like in numbers.
45 minutes, freeData & feasibility audit
We assess data quality, edge constraints and integration points, and return a written go / no-go with an architecture sketch.
About 2 weeksPilot on your site
A working model on one line, herd, robot or site, deployed on real hardware and measured against the agreed targets.
8-12 weeksScale & operate
Roll-out across sites with monitoring, retraining and handover documentation, so your team owns the system.
Ongoing, as neededFAQ
Questions we hear on the first call
How long does an edge AI pilot take?
A scoped pilot on your own data typically runs 8 to 12 weeks: two weeks for a data and feasibility audit, then iterative model development and a field deployment on one line, herd, robot or site.
Do we need new hardware?
Usually not. We act as an intelligence layer on top of the cameras, PLCs, sensors and controllers you already run. When an edge device is needed we recommend commodity hardware such as NVIDIA Jetson or industrial x86 gateways.
Can our data stay on-premise?
Yes. Our pipelines run on your infrastructure or a private cloud, and for multi-site or sensitive data we use federated learning so raw data never leaves its source.
Do you join EU-funded consortia?
Yes. We have delivered cascade-funded and Horizon Europe work as a technical partner, contributing edge AI, MLOps and integration expertise, and we co-write proposals with partners.
What is TraceLM and who is it for?
TraceLM is our platform for steel processors and traders. It reads mill test certificates and tensile-test PDFs with OCR and LLM extraction, validates the data, and issues EN 10204 3.1 and 2.2 certificates with full traceability from mill to customer. A Digital Product Passport plugin prepares products for the EU ESPR rules. It runs in production at an Athens Stock Exchange-listed steel company.
Where do you work?
We are based in Athens, Greece and work with clients and consortia across the European Union, on-site when needed and remotely by default.
Book a discovery call
Your question is not here? It will be answered on the call, in writing afterwards.
Contact
Tell us what your data should be doing
Describe the problem in a few lines. An engineer, not a sales rep, replies within one business day with next steps or honest reasons why AI is not the right tool.
- info@plaixus.com
- Based in
- Athens, Greece · working across the EU
- Hours
- Monday-Friday, 09:00-18:00 EET