Services
Senior, hands-on engineering, owned end-to-end — from architecture through to running it live. Available for project-based and long-term engagements.
LLM & agentic systems
Autonomous, multi-step agents that plan, use tools, and act — with structured, verifiable output rather than free-form text. Agent orchestration, retrieval-augmented generation, and generative-AI integration into existing products and workflows.
Document intelligence
Extracting clean, structured data from messy, inconsistent real-world documents — combining text extraction, layout detection, and vision-language models to replace slow, error-prone manual data entry.
Any model, anywhere
Models from OpenAI, Anthropic, Google, and the open-weight ecosystem — chosen per problem, not per vendor. Where data can't leave your infrastructure, open-weight models run self-hosted on your own hardware, with nothing sent to any third party.
Architecture & data platforms
The foundations underneath: systems and data architecture, full-text search over millions of records, long-lived production systems, and the integration pipelines that keep everything in sync.
Selected work
Document intelligence for a Swiss wholesale distributor
An end-to-end system extracting structured sales orders, purchase orders, and invoices from complex, inconsistent, multi-language B2B PDFs — combining text extraction, layout detection, and vision models, and cross-checking document totals against extracted line items. Commercial confidentiality ruled out external AI APIs, so open-weight LLMs run self-hosted on the company's own infrastructure. Replaced manual data entry in the company's procurement workflow.
Python · FastAPI · PostgreSQL · SGLang · vision-language models
Autonomous content platform
An AI-driven social-media system in which autonomous agents plan, generate, and schedule content end-to-end — multi-step agent orchestration, a computer-vision pipeline for media understanding, and automated scheduling.
Python · FastAPI · PydanticAI · vision models · Svelte
Search & scale for an online book retailer
An e-commerce platform built from scratch — no off-the-shelf CMS — with fast full-text and faceted search across a catalogue of over 6 million titles, fed by automated pipelines keeping pricing and availability current across millions of records. Its LLM content pipeline generates book descriptions grounded in the catalogue's structured data, with prompts assembled combinatorially so the writing stays specific and varied across hundreds of thousands of pages.
Python · PHP · PostgreSQL · Sphinx · Linux · Nginx
Sound like your problem? contact@insomniacsoft.com