AI & Machine Learning

We build AI that ships as part of a product: agents that act inside your workflows, LLM pipelines with retrieval over your documents, forecasting and scoring models, computer vision and document AI — with monitoring and an owner, never a separate science project. Our AI team works inside the same delivery process as the rest of Glanit: discovery, pilot, rollout, support. This is now mainstream work: McKinsey reports that 78% of organisations use AI in at least one business function (McKinsey, 2025), and the difference between them is not the model but whether it reached production.

• AI agents that act inside your workflows: read and write CRM/ERP records, answer customers, prepare documents, escalate to people when confidence is low.
• LLM pipelines with retrieval (RAG) over your documents, tickets and databases — grounded answers with sources.
• Forecasting and scoring: demand, churn, risk, pricing, lead quality.
• Computer vision: quality control, document and image classification, OCR.
• Document AI: invoices, contracts, forms — extraction, validation, posting to the ERP.
• NLP: classification and routing of feedback, emails and tickets in many languages.

For most business cases the answer is retrieval (RAG) first: it grounds answers in your current documents with sources, updates instantly when the documents change and needs no training run. Fine-tuning is worth it when you need a specific style, format or narrow skill at lower cost per call. The two combine well; retrieval-augmented generation was introduced by Lewis et al. (2020) precisely to give language models fresh, citable knowledge without retraining.

An AI project runs in four stages — discovery, pilot, rollout, support — and every stage ends with a number you can check. The reason for the discipline is simple: Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, unclear business value and escalating costs (Gartner, 2024). A pilot that is measured against your real cases from week one does not become one of them.

Discovery (1–2 weeks). We map the process, collect a sample of real data and agree on the metric that decides success — hours saved, error rate, conversion, response time.

Pilot (4–8 weeks). A working system on a narrow slice with an evaluation set built from your real cases. You get a number, not a demo.

Rollout. Integration with your systems, guardrails, human-in-the-loop where actions are irreversible, monitoring of quality, cost and drift.

Support. Evaluation on every change, model updates, an improvements backlog — under an SLA.

The stack is Python first: PyTorch, scikit-learn, LangChain/LlamaIndex or plain SDKs; OpenAI, Anthropic, Google and open-weights models (Llama, Mistral, Qwen) — self-hosted when data must stay inside your perimeter. Vector search runs in pgvector, Qdrant or Elasticsearch; serving on Kubernetes or serverless; tracing with OpenTelemetry/Langfuse. We pick the smallest model that reliably hits the target, because prices keep falling: Stanford's AI Index estimates that the cost of running a model at GPT-3.5 level fell more than 280-fold between November 2022 and October 2024 (Stanford HAI, 2025) — a good reason not to lock the architecture to one vendor.

The proof is in the case studies and in the numbers they report. See our case studies: retail assistants on RAG, invoice Document AI, catalogue deduplication, customer-feedback NLP, and our own AI products. We also explain how AI agents work in production, and every AI system we ship is built with our AI security practice in mind.