What we do

1. 2.1 Systems that replace spreadsheets

   The sign a spreadsheet has outlived itself is not its size. It is the moment when only one person is willing to edit it.

   We build the applications that take that work over: access scoped by role, an audit log showing who did what and when, processes that run on a schedule instead of from memory, and reports that exist rather than get assembled.

   FastAPI and Python at the back, React and Next.js at the front, PostgreSQL underneath. We use these because we have maintained them in production, not because they are current.

   Our own rental management platform — four portals, 131 endpoints, 858 tests, in continuous delivery since July 2026. An internal ERP with six modules that we use daily. Case study: Managing Rentals

2. 2.2 Compliance that tracks itself

   Intrastat, NAP filings, workplace-safety regulation. Three obligations with monthly deadlines, exact formats, and penalties that do not care whether somebody forgot.

   We build the tools that carry that weight. Declarations generated in the required format instead of assembled in Excel. Validation that shows the problem row before submission rather than after rejection. Statutory safety deadlines that announce themselves in advance.

   We know this material because we administer it for our own company.

   A desktop Intrastat tool delivered to a Bulgarian trading company, with more test code than application code. A workplace-safety module in our own internal system. Case study: Intrastat tool

3. 2.3 Retrieval and agents, built for real inputs

   Most AI demonstrations work because the input was chosen to make them work. Production inputs are scanned, rotated, multilingual, badly structured, and occasionally wrong.

   We build the systems that survive that: retrieval you can tune rather than hope about, extraction that escalates its strategy when quality drops, answers that carry their sources, and behaviour you can trace when something goes wrong at three in the afternoon.

   Python throughout — FastAPI, LangGraph, vector stores, whichever model provider the problem actually needs.

   Delivered under NDA. We can walk through architecture and decisions in a call. Our own products in the same field: DocumentInsight · ArticleCraft AI

4. 2.4 Knowing whether it works

   Most teams ship AI features on spot-checks and optimism. Then something regresses quietly, and nobody can say when it started or what caused it.

   We build the measurement layer: scoring rubrics for your actual task, offline metrics that run in CI, regression suites that fail loudly, groundedness checks, and a human review loop small enough that somebody will actually maintain it.

   This is our research area, not a service line we added because the market asked. Maria's doctorate is in AI, and her layered evaluation framework for multi-agent LLM systems is under peer review.

   Four evaluation layers, applied to a production multi-agent document pipeline. Paper under review.