An AI-ready software company
We build business software and the AI systems around it: agents, harnesses, retrieval over your documents, and Keepwork, our own app builder. From Sofia, for clients abroad.
- Where
- Sofia, Bulgaria. Remote or on site, in English and Bulgarian.
- Models
- Anthropic, DeepSeek, Gemini and OpenRouter in the cloud; Ollama and LM Studio on your own hardware.
- AI stack
- Agents, harnesses, evaluation, sandboxed runtimes, pgvector retrieval, OCR pipelines.
- Software stack
- .NET, Go, React, Next.js, Flutter, PostgreSQL, Azure, AWS, Kubernetes.
What we build
Four kinds of work. Each one has been done on our own products before we offer it to you.
AI agents and harnesses
A harness around the model that plans, calls tools, checks its own work and records every step, so an agent can do real work in your systems instead of chatting about it.
Retrieval over your documents
Your documents and data behind a model, with every answer traced to its source and every source checked before it is used. Built for the questions your team keeps asking.
Agent kits for engineering teams
Skills, sub-agents with roles, guardrails as code and review gates, set up in your repository, so your coding agents ship verified work rather than claim it is done.
Product engineering and platforms
The software itself: web and mobile products, services, identity and access, cloud. Eighteen years of it, in energy, industry, fintech and healthcare.
Now building
A front-desk AI assistant for organisations that cannot afford a wrong answer.
With a partner company we are building an assistant that answers calls and chats in Bulgarian and English from the client's own published knowledge, hands off to a person with a summary when it cannot help, and keeps every client's data apart at the database level. It runs in the cloud or on the client's own hardware, with the model providers the client permits, and it always says it is an AI.
It is in development and not yet in customers' hands. The details are the client's, so I share them on a call.
Keepwork, our own product
A team describes an app in plain language. The model writes it, and Keepwork's runtime checks it and runs it on a server you own. It is the harness we built first, and everything we learned building it goes into client work.
It is not released yet. I install it myself, so the first conversation is a call.
Your server
Installed on your own server or NAS. Your data never leaves it.
Your choice of model
Anthropic, DeepSeek or OpenRouter, or a local model through Ollama or LM Studio.
Sandboxed by design
Model-written code runs with no authority of its own. The boundary holds when the model gets it wrong.
No per-seat fees
One installation, however many people use it.
Writing
How the work actually goes, with the measurements and the diagrams. The proof for the claims above.
, 8 min read
If your model writes code, run it in a secure sandbox. Here are two ways to build one.
If you want a model to generate code, you have to secure the place where that code runs. Prompts, output filters and "please don't" instructions are not a boundary.
, 12 min read
Claude designs, DeepSeek builds, and they only talk through two files
Parts 1 to 3 covered why I rewrote my guesthouse back office, what I rewrote it into, and the guardrails that refuse unverified work.
, 9 min read
Guardrails as code: how I let a fast, cheap model write 19,000 lines of Go
In Part 2 I described rewriting my guesthouse back office in Go. DeepSeek V4.1 Flash wrote most of the code, running in DeepSeek Harness.
Selected work
Industries I have built systems for. The clients stay unnamed; the problems and the technology are real.
Clinical information platform with retrieval
A Bulgarian-language reference for healthcare professionals: a registry of about thirteen thousand medicines, scanned regulator documents digitised with OCR, and question answering that cites its sources and never gives clinical advice.
Next.js, Firebase, Gemini embeddings and generation, pgvector-style retrieval, Python OCR
Energy sector enterprise software
Cloud-native .NET services, sign-in across several identity providers, and a role model that grew from one role to configurable multi-application roles, designed to be AI-ready.
.NET, Entra ID, CIAM, Keycloak, Kubernetes, AWS
Cement plant operations platform
Data from many sensors, collected and processed for monitoring and operational insight, with a React front end on top.
React, .NET 8, RabbitMQ, gRPC, Databricks, Azure
Fintech back office, South America
Existing payment providers integrated behind one product, with internal tools and scalable financial workflows.
Next.js, .NET 8, MongoDB, CQRS
Medical software from scratch
Video and real-time features for a new product, built with a team of four to five that I led. Later, a plan to move a legacy Angular platform to Next.js with module federation and OpenAPI contracts.
React, TypeScript, WebRTC, WebSockets, Next.js
Digital asset management for music
A platform that delivers music from artists to stores, built as microservices, with performance work on an 800-million-record workload.
.NET, Azure, microservices
How we work
Short cycles, a named person on the code, and a plain account of what is going well and what is not.
A call
Thirty minutes. You tell me what you need built and what is in the way.
A proposal
A written scope and plan, with the risks named, so you can decide.
Delivery in short cycles
Working software every cycle, with the AI parts measured, so you see real progress and can change direction early.
Handover and support
Your team can run and extend it. I stay available for the questions that come after.
Tell me what you need built
Thirty minutes, no preparation needed. You describe the problem and I tell you honestly whether and how we can help, and where AI belongs in it.
Or write to ivaylo@blazhev.software.