Practical ways to make everyday work easier.

Technical contribution

Konect turns contact centre conversations into information for service quality assessment, structured extraction and call metrics.

Transcription with speech engines such as Deepgram and Whisper, structured extraction and conversation analysis with language models. Search over calls and asynchronous processing, with model selection informed by privacy and performance requirements.

Aritz contributes to architecture, model selection and the integration of speech processing and LLM analysis. This AI project is part of a collaboration with Irontec since 2025 that also includes software development.

Technical contribution

Qamarero wanted to introduce phone agents into its hospitality SaaS, connected to restaurants’ daily operations.

Aritz implemented the phone agents and integration tools that check availability and connect bookings with the SaaS CRM, tables and opening hours. Several months of collaboration with Qamarero’s team through production delivery.

A complete conversation and booking flow connected to the software the restaurant uses.

Technical contribution

A property management workflow with hundreds of thousands of documents per month needed classification into 15 categories and 77 subcategories. At this volume, per-document cost, latency and ranking quality guided the architecture.

We evaluated generative LLMs, centroid methods, hybrid routing and CSLS hubness correction. Geometric analysis showed that centroid similarities compressed useful information. The final evaluated design combines four bagged logistic-regression variants over 154 semantic similarities with a classifier over the complete 1,024-dimensional embedding. A geometric mean merges their probabilities to rank 77 subcategories. Classification runs on CPU, without generative LLM calls or an extra embedding per document.

The evolution reduced cost and improved classification compared with the initial all-LLM approach. Mathematical advice helped guide the geometric analysis. The selected ensemble also exceeded the tested generative rerankers, retaining local, reproducible classification after document extraction and embedding.

Final experiment report, stored results and implementation in LAE_CATEGORIZACION (lae-parser). These are development evaluation results; the monthly operational volume is separate from the test sample.

Technical contribution

Child protection teams need to coordinate case files, intervention notes, documents and daily follow-up over highly sensitive information. AI must fit the professional workflow and respect each user’s permissions.

Platform architecture and development: case files, notes, documents, activities, calendars and reports, with roles and security logs. FastAPI, PostgreSQL and asynchronous ARQ/Redis jobs. The assistant combines Typesense hybrid search and local Ollama models, cites original notes and manages a limited context budget. Draft generation includes iterative synthesis, grounding checks and fallbacks.

As co-founder and CTO, Aritz leads architecture and development and led and implemented compliance with the Spanish ENS at Medium category. The professional team reviews answers and reports.

Co-founded platform. ENS Medium compliance confirmed in September 2026.

Technical contribution

An expert report needs consistent dates, injuries and assessments. Source documents and professional judgment must guide every section.

A 12-node LangGraph workflow organised by section, extracting validated JSON from complete source documents. Deterministic date and consistency checks work alongside an LLM judge evaluating each section against a reference report. Langfuse traces support diagnosis; model routing by task is guided by evaluation.

Evaluation enabled model cost adjustments while preserving measured quality. The professional reviews the final report.

Technical contribution

A law firm’s information is spread across case files, email, documents and calendars. These tools need connecting while preserving professional control.

A management platform covering clients, case files, Gmail, Drive, Calendar, tasks and document generation. AI suggests tasks and attachment classification; a lawyer reviews and accepts before execution.

Workflow integration and AI assistance with an explicit human approval step.

Technical contribution

Analysing court decisions means converting PDFs into consistent data and handling ambiguous cases explicitly.

Document extraction, classification into legal categories and a second pass for ambiguous cases. JSON output validated against data schemas, with accuracy and consumption reports.

A system documented in code combining document processing, structural validation and evaluation.

Technical contribution

Case tracking needs to gather information from legal software and check relevant data before taking action.

Integration with legal software, structured extraction of status, dates and registration numbers, a human review panel, controlled notifications and auditable PDF records.

Automation with traceability and review of the data involved in each action.

Technical contribution

A long report needs relevant sources for each section and consistency across its narrative.

The kompletai-rag-langgraph pipeline retrieves citations, curates duplicates, drafts sequentially and runs a correction pass to flag inconsistencies and potential hallucinations. Local Ollama models are configured by task.

Technical development with explicit retrieval, drafting and review stages. Reports require professional validation.

Technical development: kompletai-rag-langgraph.