Computer vision, LLM and MLOps, plus document analysis on local AI

BeGiga carries out artificial intelligence projects for companies, organisations and R&D departments, spanning computer vision, systems based on large language models, and the infrastructure that keeps those models alive in production. Work can proceed end to end, or the BeGiga team can slot into an existing team and handle one selected stage. Scope is always shaped by the problem at hand, the data at hand and the client's infrastructure, while a feasibility check before kickoff forms part of the system design and technology selection service.

Computer vision

BeGiga delivers projects and partners on R&D efforts in computer vision. Everything begins with dataset engineering: deciding which photos and videos must be gathered, drafting labelling guidelines, verifying annotation quality and using augmentation that makes a model resilient to situations that seldom show up in training data. Built on that groundwork are custom models for detecting, segmenting, classifying and tracking objects across images and video, along with solutions working on 3D data.

Each model is built around the actual data it will face once deployed: weak lighting, motion blur, odd camera angles and noise absent from lab conditions. Image and video processing pipelines link the model to surrounding stages, including frame preparation, filtering of results and forwarding them into the client's systems. Where a model must operate on a phone or an embedded device, it gets quantised and tuned for the target hardware, so that on-device inference stays within memory, response time and power constraints.

LLM systems, RAG and agentic AI

BeGiga designs and rolls out systems based on large language models, where quality hinges first and foremost on data preparation. This work involves pulling content out of unstructured documents, including PDFs, Word files, presentations and scans, and converting it into structured datasets that are cleaned, split into chunks and enriched with metadata. Those chunks are then converted into embeddings and indexed in vector databases like Qdrant, Milvus, pgvector or FAISS, selected according to project scale and the environment where the system will operate.

Running above that index are semantic search, keyword search and hybrid search that merges the two, plus metadata filtering and reranking. RAG systems draw their answers from the company's documents and cite the source behind each answer, which cuts down hallucinations and lets anyone verify the origin of the information. Models may be accessed via commercial providers' APIs or operated locally as open-source models, with fine-tuning on the client's data applied when the task demands it.

Agentic pipelines take things a step further: they map out the steps of a task, invoke tools and functions of other systems, query databases and merge outputs from multiple models. Integrations may operate via the MCP protocol, which makes company systems' data and functions available to models in a controlled fashion. Each such system receives guardrails against prompt injection and unwanted output, action and cost limits, logging, and checkpoints where a person signs off on the decision. Quality gets measured on test sets created for the specific use case, separately for retrieval and for answers, and once in production, drops in quality and call costs are tracked.

A simpler, public-facing variant, namely a chatbot built on website content, appears in the article on website chatbots. One example of AI embedded in software a company already runs is the PrestaShop assistant and MCP server, discussed in should you upgrade PrestaShop to 9.2.

MLOps and data engineering

BeGiga establishes MLOps practices that carry a model from prototype into production and hold it there without manually repeating identical steps. This encompasses experiment tracking and a model registry in MLflow, versioning of data and models, containerisation, automated testing and releases via CI/CD, model serving, plus monitoring of model performance and data drift. Once data drifts far enough that a model's performance degrades, a retraining pipeline rebuilds it and benchmarks it against the earlier version before the swap takes place.

A model requires a constant flow of data, and that is what ETL and ELT pipelines supply. Under ETL, data gets extracted from source systems, transformed and loaded into a data warehouse only afterwards. Under ELT, raw data arrives at its destination first, for instance a lakehouse, and transformation happens there, which suits large and frequently shifting datasets. Orchestration in Apache Airflow manages scheduling and task order, executing steps in a defined sequence and retrying them following failures.

At large data volumes BeGiga relies on Apache Spark, an engine for distributed batch and streaming data processing, and with the Databricks platform, including the Unity Catalog data catalogue, MLflow, model serving and vector search. Data, models and their history then live in a single place, governed by access control and with every step reproducible. Smaller projects follow the same principles using lighter tools matched to data volume and budget.

A tailored document analysis stack with local AI

We also build and deploy document archiving and analysis stacks fitted to a particular company, preparing its documents for further processing by AI. Rather than a single boxed product, it consists of proven open-source tools and add-ons configured around the company's document types, user count and working style, connected to email, scanners and network folders and to local AI models.

Each document, scanned or born digital, passes through text recognition (OCR), making it searchable by any word it contains. The system itself assigns the document type, correspondent and tags, learning from previous descriptions, extracts attachments from mailboxes while ignoring unneeded files, and labels paper originals with archive serial numbers (ASN) as barcodes or QR codes. That ties each original unambiguously to its digital copy and makes its place in the paper archive easy to find. Permissions, meaning who can view and edit which documents, are settled before configuration begins.

This sort of deployment suits companies that lack an extensive document management system so far: service offices, small manufacturers and traders, plus teams operating in the field. In places where documents already flow through an ERP, CRM or corporate suite, a different route often works better, and we say so openly before any work starts. To see why a structured document collection is the starting point for an internal knowledge base built on LLM technology, read documents first, then AI.

Privacy, local models and hardware

Company documents need not travel to an external cloud. The stack may operate on a computer or server in the office, over a private network, with outside access via a tunnel or VPN, or, where convenience takes priority, on a private VPS with a private network. In each case the files stay out of public reach, and the configuration is adjusted to the GDPR requirements relevant to the company.

The same principle covers AI features. Rather than shipping documents to an outside provider, the company may rely on local language models operating on its own hardware, with BeGiga designing that server and selecting the components, including a graphics card suited to the chosen model. For which hardware fits which model, see running a local LLM in a company: which hardware for which model. The well-organised document collection then turns into the foundation for internal assistants answering questions about document content and pointing to the source, for meaning-based search, summaries and inconsistency checks across contracts, invoices and protocols. These are the very RAG systems covered in the language model section, free from dependence on AI providers' pricing and terms.