iotain plugs AI and autonomous agents into your existing systems, fine-tunes and secures models with SFT, RLHF, RPO and red-teaming, builds the rubrics that make them accurate, and turns hours of shop-floor video into labelled, searchable time.
We don't replace the platforms you rely on. We make them smarter, keep your data where it belongs, and hand you models you can measure.
We connect language models, agents and automations to the tools your teams already use, with no rip-and-replace.
AI agents that plan, use your tools and complete multi-step work on their own, with guardrails and human approval where it matters.
We adapt open and hosted models to your domain, tone and tasks, then align them to the answers your experts prefer.
We define what "good" looks like for your use case, measure your model against it, and keep training until it hits your accuracy targets.
We attack your AI the way a real adversary would, then fix what we find, so your assistants and models are safe to put in front of customers.
We build, train and deploy your own models with PyTorch and TensorFlow, from classic machine learning to deep learning and computer vision.
Every frame marked with the product in view and whether the moment was productive, idle or non-productive, ready for analytics or for training your own vision model.
Our engineers build the product around the model: web and mobile apps, SaaS platforms, APIs and back-ends, ready for production.
An AI agent does more than answer. It plans the steps, calls your systems, checks its own work and asks a person before anything important happens.
A request arrives by chat, email, ticket or schedule.
The agent plans the steps and picks the right tools.
It reads and writes in your CRM, ERP, inbox or database.
Results are checked, and high-risk steps wait for a person.
Every action is logged for audit, cost and quality review.
We label your CCTV and process footage frame by frame, combining model pre-labelling with human review, so every minute is tied to a product and an activity state.
| Start | End | Zone | Product | State | Conf. |
|---|---|---|---|---|---|
| 14:02:11 | 14:09:40 | Station 1 | SKU-2231 | productive | 0.97 |
| 14:09:40 | 14:12:58 | Cart B | — | idle | 0.93 |
| 14:12:58 | 14:21:05 | Station 1 | SKU-2231 | productive | 0.96 |
| 14:21:05 | 14:24:17 | Station 3 | SKU-1187 | non-productive | 0.89 |
| 14:24:17 | 14:36:02 | Station 2 | SKU-1187 | productive | 0.95 |
Sample export · one row per labelled segment · CSV, JSON, COCO or YOLO formats
We teach the model your domain with supervised fine-tuning, then align it with human preferences using RLHF or RPO, and prove the gain with rubric scores.
Collect your documents, chats and tickets; deduplicate and redact personal data.
Instruction and response pairs written or reviewed by your domain experts.
Annotators rank model answers side by side against your rubric.
Reward modelling with RLHF, or direct preference optimisation with RPO.
Rubric scoring on a held-out test set, then deployment in your cloud.
A rubric turns "it seems better" into a score. We write rubrics with your team, grade outputs with experts and AI judges, and feed the results back into training.
| Criterion | Weight | Meets (2) | Partial (1) |
|---|---|---|---|
| Factual accuracy | 35% | Every claim matches the source material | Minor error that doesn't change the outcome |
| Follows instructions | 25% | Answers the exact question in the required format | Right answer, wrong format |
| Policy & safety | 20% | Stays within your policies and tone | Correct but off-brand tone |
| Reasoning shown | 20% | Cites the source it relied on | Generic reasoning |
Every assistant you put online can be tricked, leak data or be pushed off-policy. We test against the OWASP Top 10 for LLM applications and fix what we find.
| Threat | What we test | How we fix it |
|---|---|---|
| Prompt injection | Direct and hidden instructions in user input, documents and web pages | Input isolation, instruction hierarchy, output checks |
| Data leakage | PII, secrets and system prompts in answers | Redaction, access scoping, response filtering |
| Jailbreaks | Role-play, encoding and multi-turn attacks | Safety tuning, guardrail models, monitoring |
| Excessive agency | Agents calling tools they shouldn't | Least-privilege tools, human approval steps |
| Data poisoning | Tampered training or retrieval data | Dataset provenance checks and anomaly scans |
Customer and staff assistants grounded in your own content.
Search and answers across manuals, contracts and SOPs.
Transcription, voice agents and call analytics.
Invoices, orders and tickets handled end to end.
Lead scoring, call summaries and next-best actions.
Forecasting, anomaly detection and natural-language reports.
EU AI Act readiness, policies, model cards and audit trails.
Product search, recommendations and catalogue enrichment.
We pick the right framework for the job, from PyTorch and TensorFlow for custom models to LangGraph for agents, and we work inside the cloud you already use.
A free session to map where AI saves time or money in your operation.
We check what data you have, its quality, and what can be used safely.
A working proof of concept on your real data within weeks.
Fine-tuning, annotation and rubric scoring until targets are met.
Connected to your ERP, CRM, cameras or website with access control.
Ongoing quality tracking, retraining and support.
One use case, one dataset, a measured result you can take to the board.
A defined build such as an ERP assistant, an annotation programme or a fine-tuned model, delivered to production.
Engineers and annotators working as an extension of your team.
iotain integrates AI into existing businesses, fine-tunes large language models with SFT, RLHF and RPO, designs learning rubrics to evaluate and train client models, and annotates video to label products and productive, idle or non-productive time. We also build chatbots, AI agents, knowledge search and process automation.
Supervised fine-tuning (SFT) teaches a model your domain from example questions and answers. Reinforcement learning from human feedback (RLHF) trains a reward model on human rankings and optimises the model against it. RPO is a preference-optimisation method that learns directly from ranked answer pairs, which is often simpler and cheaper to run than full RLHF.
We combine model pre-labelling with human review to mark every segment of your footage with the product in view and an activity state: productive, idle or non-productive. You receive frame-accurate timelines in CSV or JSON, plus training datasets in COCO or YOLO format for your own computer-vision models.
Agentic AI means AI agents that can plan and carry out multi-step tasks using your tools, such as updating your CRM, processing invoices or answering support tickets end to end. We build agents with clear permissions, human approval for high-risk steps and full audit logs.
We train and deploy models with PyTorch, TensorFlow, Keras, scikit-learn and Hugging Face, build agents and RAG systems with LangChain, LangGraph, LlamaIndex and MCP, and run them on AWS, Azure or Google Cloud with MLflow, Docker and Kubernetes. We work with OpenAI, Claude, Gemini, Llama, Mistral and other models.
We red-team your model against the OWASP Top 10 for LLM applications, testing prompt injection, jailbreaks, data leakage and unsafe tool use. Then we add guardrails, input and output filtering, least-privilege tool access and monitoring, and retest until the issues are closed.
Yes. We connect AI assistants and agents to the systems you already use, such as SAP, Microsoft Dynamics, Odoo, Salesforce, HubSpot or a custom platform, without replacing them.
Your data stays under your control. We can train and deploy models in your own cloud or on-premise, redact personal data before training, and work with OpenAI, Azure, Claude, Gemini or fully private open-source models.
Book a free AI discovery workshop. We map where AI can save time or money in your business and propose a pilot with a measurable result.
Tell us about your systems and your goal, and we'll set up a call.