Services
Five ways to work with us. Each one is led by a senior engineer who stays on the code, and each one has been done on our own products first.
AI agents and harnesses
A model on its own answers questions. A harness around it does work: it plans, calls your tools, checks the result, retries within limits and records every step.
What you get
- A harness built for your task: planning, tool calls into your systems, checks after every step, bounded retries, and a recording you can replay.
- Work split by cost and strength: a strong model designs and reviews, a fast and cheap one executes, with written contracts between them.
- Model routing: Anthropic, DeepSeek, Gemini or OpenRouter in the cloud, or Ollama and LM Studio on your own hardware, chosen per task and switchable later.
- Evaluation that gates changes: a benchmark of real tasks, run before anything ships, so a prompt or model change cannot quietly make things worse.
- Sandboxed execution for model-written code, with no authority inside the sandbox, so the boundary holds when the model gets it wrong.
When it fits
You want a model to carry out work in your systems, such as processing requests, drafting and checking documents or building and running small tools, and you want to know how often it gets it right.
Retrieval over your documents
Your documents, registers and records behind a model, so people get answers with the source attached instead of a confident guess.
What you get
- Ingestion that fits the material: OCR for scanned documents, splitting along the structure the documents already have, validation and a confidence score for every piece before it enters the index.
- Chunking that respects sentences and sections, embeddings in PostgreSQL with pgvector or in the store you already run, and retrieval tuned on your real questions.
- Citation-only answers where it matters: the model may only state what a retrieved source says, and shows the source. In regulated fields it never adds advice of its own.
- Permissions carried through: a person only gets answers from documents they may see, enforced in the database rather than in the prompt.
- Usage limits, streaming replies, and a review queue where staff correct the knowledge instead of the model.
When it fits
Your team keeps asking the same questions of the same documents, or your customers do, and a wrong answer costs you something.
Agent kits for engineering teams
Your developers already use coding agents. A kit makes the environment refuse unverified work, so "done" means done.
What you get
- Skills: written workflows the agent loads when it needs them, such as plan, implement, research or review before a pull request.
- Sub-agents with roles: implementer, test author, bug fixer, and an adversarial reviewer whose only job is to break the change.
- Guardrails as code: rules that refuse an action outright, with the reason in the refusal. Scopes per agent, protected paths, secrets kept out of reach.
- Hooks and gates: checks after every edit, tests and a definition of done before a task may be called finished, a limited number of retries, then a human.
- Set up in your repository for Claude Code or the harness you use, with the team shown how to work with it.
When it fits
You have a team using Claude Code, Cursor or similar, and you want speed without reading every diff yourself.
Product engineering and platforms
The software the AI parts live in. Built end to end by a senior engineer who has done it for eighteen years and still writes the code.
What you get
- Web with React and Next.js, mobile with Flutter, services in .NET or Go, PostgreSQL underneath.
- Microservices and event-driven systems with RabbitMQ, gRPC and MassTransit, on Azure or AWS with Kubernetes and GitOps.
- Identity and access: sign-in across providers with Entra ID, Keycloak and CIAM, roles that can grow with the product.
- Architecture reviews, OpenAPI contracts and staged modernisation plans, and a lead for a small team when you need one.
When it fits
You have a product to start, a system that has outgrown its first version, or a legacy platform that must move without stopping the business.
Keepwork installations
Our own AI app builder, installed on your server and set up with your first app.
What you get
- A thirty-minute call, then one day on your server: install, backups and your first app built together.
- A week of real use, with fixes, and a handover to your team.
- A yearly care plan: updates, backups checked every month, and support.
- No per-seat fees. For the terms, book a call.
When it fits
Your office keeps asking for small internal apps, such as bookings, approvals or trackers, and you want them on your own server without hiring a developer.
Not sure which one you need?
Book a call and describe the problem. I will tell you which of these fits, where AI belongs in it, or that none of them does.
Or write to ivaylo@blazhev.software.