About

About

I build LLM agents and document processing systems that run in production.

As ML Architect at Graip.AI, I architect the full stack behind these systems: data pipelines, analytics and predictive models, and the infrastructure for training, serving and monitoring models at scale.

My work is grounded in research. My doctoral studies in NLP at the University of Latvia, supervised by Prof. Guntis Barzdiņš, focused on enriching model inputs with LLM-generated context, the same idea that now powers retrieval in Agent Labs. Now I study when a task truly needs a large model and when a smaller one will do.

On this blog I write about production LLM agents, RAG, document AI and NLP research.

Experience

Graip.AI · ML Architect · 2023–present

  • Agent Labs: a platform for LLM agents that answer questions over company documents, query databases and act in business systems like SAP, with human approval where it matters.
  • Document processing (IDP): pipelines on the Graip.AI platform that extract structured data from invoices, contracts and forms, combining LLMs and fine-tuned models.
  • ML workflows and infrastructure: data pipelines, predictive models, and services for training, serving and monitoring models (FastAPI, Kubernetes, MLflow).

University of Latvia · Senior Researcher and Consultant · 2023–present

  • Research: efficient use of large language models for information extraction; methods that enrich model inputs with LLM-generated context, later applied in production retrieval at Graip.AI.
  • Teaching: a YouTube mini-course on training neural networks from scratch, and articles on deep learning for NLP.

If Insurance · ML Architect · 2020–2023

  • End-to-end ML workflows: from data ingestion and preparation to training, evaluation and deployment, with MLflow for experiment tracking and reproducible releases.
  • Deep learning and NLP: fine-tuning pre-trained PyTorch models; text processing and feature extraction with spaCy and NLTK.
  • Data engineering and analysis: data transformation and ingestion with pandas, PySpark and SQL; visualizations that informed business decisions.

Scandiweb · Senior Software Engineer · 2017–2020

  • Backend development and data engineering in Python: APIs, ETL pipelines and deploying ML models, delivered for clients such as Pins (airBaltic) and sportland.lv.

Earlier · Software Engineer · 2014–2017

  • Web applications and ERP/CRM systems at DYNINNO, EVO Group and as a freelancer.

Education

  • University of Latvia: PhD studies in Computer Science (2020–present, thesis under review); MSc in Computer Science (2018–2020)
  • Riga Technical University: Professional Bachelor in Mechatronics (2012–2018), with an Erasmus+ semester at TU Dresden

Publications

With Guntis Barzdiņš. The development workflow behind the shared-task papers is described in this post.

  • 2026 · Engineering a Production RAG-Based Agent Platform. Baltic DB&IS 2026, CEUR-WS Vol. 4211. PDF
  • 2025 · RIGA at SMM4H-HeaRD 2025: Context-enriched classification pipeline. SMM4H-HeaRD workshop at ICWSM 2025. Paper · Code
  • 2024 · RIGA at SMM4H-2024 Task 1: Enhancing ADE discovery with GPT-4. SMM4H workshop at ACL 2024. Paper · Code
  • 2023 · RIGA at SemEval-2023 Task 2: NER enhanced with GPT-3. SemEval-2023 at ACL 2023. Paper · Code
  • 2022 · RIGA at SemEval-2022 Task 1: Scaling recurrent neural networks for CODWOE dictionary modeling (with Gus Strazds). SemEval-2022 at NAACL 2022. Paper · Code

Work with me

I consult on LLM agents, document AI, data engineering, data science and ML service architecture. What you get:

  • Agents you can trust in production: orchestration, tool integrations and human approval designed so agents can act in your systems safely.
  • Better answers from your documents: retrieval tuned with hybrid search, re-ranking and contextual enrichment.
  • Less manual data entry: extraction pipelines that turn invoices, contracts and forms into structured data.
  • Lower model costs: find where a smaller fine-tuned model can replace large-model calls without losing quality.
  • Confidence before release: evaluation sets and regression checks that show whether a change actually helps.
  • A faster path from data to production: end-to-end ML workflows and service architecture for training, serving and monitoring.
  • Data you can build on: data pipelines and analysis that turn raw data into models and business decisions.

I’m also open to research collaborations, talks and team workshops. Get in touch by email or LinkedIn, both in the sidebar.