Full-Stack Software · AI Agent Systems
Software that ships.
Agents you can trust.
We design and build production software with AI agents inside — RAG knowledge systems, agent teams with human-approval gates, and deep ERP/CRM integration. From architecture to daily production use.
30 minutes with the engineer who would build it — not a sales team
Most AI initiatives die between demo and production.
The gap isn't the model — it's engineering: retrieval quality, approval gates, integration, and verification.
Demos that never reach production
A proof-of-concept impresses in a meeting, then dies. Without retrieval quality gates, evaluation, and real error handling, it was never going to survive contact with production data.
Agents nobody can trust
An agent that acts without approval gates, cites nothing, and never refuses is a liability, not an asset. Trust is an engineering discipline: human sign-off on decisions, citations on claims, refusal on low confidence.
Knowledge locked in silos
The answers exist — in SOPs, the ERP, the CRM, a decade of email — but nothing that needs them can reach them. Retrieval and integration are where the real value sits.
What We Build
Six disciplines, one team, production standards.
Agentic Platforms & Agent Teams
Multi-agent systems that run real workflows end to end — every consequential action behind a human-approval gate. One agent, one task, full auditability.
RAG & Knowledge Systems
Retrieval-augmented generation your team can rely on: reranking, mandatory citations, refusal on low confidence — over text and image corpora at 5,000+ document scale.
MCP & Systems Integration
Model Context Protocol servers and deep integrations into the systems you already run — Microsoft Business Central, HubSpot, calendars, telephony — with safe, scoped access.
Full-Stack Product Engineering
From data model to shipped interface: React and TypeScript front ends, typed APIs, PostgreSQL and MySQL backends — engineered, tested, documented, handed over.
Applied Machine Learning
Not everything is an LLM problem. Neural-network modelling for constrained and embedded problems — from sensor error-compensation to price prediction.
Automation & Voice
Orchestrated automation for the unglamorous work: document and email processing, outreach pipelines, AI voice intake wired live into your CRM and calendar.
Recent Work
Client engagements are shown anonymized under contract. Two of these products are our own.
Enterprise knowledge systems · US
Two production RAG systems for a multi-billion-dollar US electronics manufacturer
An engineering-SOP retrieval system and a multimodal process-engineering system (text + engineering diagrams), serving the engineering organization daily. Trust was the design requirement: cross-encoder reranking, mandatory citations on every claim, and refusal when retrieval confidence is low.
Industrial ML · Automotive components
Sensor error-compensation on embedded systems
Neural-network models compensating sensor error on embedded hardware for a global automotive-components group — machine learning under real physical constraints, shipped to R&D.
Full-stack SaaS · Denmark
Booking & credit platform for a staffing marketplace
A transactional web application — credit engine, booking flows, admin — built with typed APIs end to end, then verified by a scripted multi-persona UAT program: three waves, 10+ defects found and fixed before handover to the client's admin team.
Vertical AI product · Legal
EstateOps — AI acquisition platform for estate-planning firms
Our own product: a domain RAG brain built on a 494-page legal corpus, an AI voice intake agent writing live into the CRM and booking real calendar slots, and an ROI calculator that opens the conversation.
E-commerce automation
ScalePod — multi-agent product pipeline
Our own product: from a product brief to published listings across Printify, Etsy, Shopify and Pinterest — pattern, mockups, copy and publishing all agent-orchestrated. 4 hours of manual work per product became 12 minutes; 97 products shipped this way.
Agentic ERP/CRM · Denmark
Supplier-pricing & quoting platform for an industrial trading company
An agentic pricing platform integrating Microsoft Business Central and HubSpot: supplier price collection, quote generation and order flow — every price and quote behind a human-approval gate, delivered inside the client's own Azure tenant.
Named references available on request.
How We Deliver
Architecture first. Working software every cycle. Verified before handover.
Architecture First
Before code: scope, data map, integration points, security model and risk register — written down, reviewed with your team, agreed.
Iterative Build
Working software in short cycles, demonstrated on real data. Human-approval gates are designed in from the first line, not bolted on.
Verify & Hand Over
Scripted acceptance tests, multi-persona UAT where it matters, documentation and a clean handover — your team owns and operates the result.
Ways to Work With Us
Every engagement is scoped and quoted after a technical conversation — no packaged tiers.
Project Delivery
A defined system, built and handed over
- Written scope with acceptance tests per deliverable
- Milestone-based delivery, demos on real data
- Staging environment + scripted UAT
- Documentation, handover and warranty
Retained Engineering
Ongoing build-and-operate capacity
- Reserved senior capacity, sprint-planned
- You own the backlog and priorities
- Continuous delivery into your environment
- Scale up or down as phases change
Architecture Consulting
Senior review before you commit
- AI feasibility & architecture reviews
- RAG and agent-system audits
- Integration and data-security design
- Build-vs-buy recommendations you can act on
You own the source, the infrastructure and the data — everything is built in your environment, on your accounts.
Who You'll Work With
Every project is architected — and largely built — by the founder.
Gergely Racz
Founder & Principal Engineer
Gergely started in machine-learning research and spent a decade taking AI from papers into production — neural networks on embedded hardware, then retrieval systems, and now agentic platforms that run daily inside real businesses. He architects every engagement personally and stays hands-on through delivery: the person you scope the system with is the person who builds it.
- PhD in Engineering, University of Cambridge
- 10+ years in AI — from machine-learning research to production systems
- 3 years dedicated to LLM-based software and agent systems
- Architect of RAG systems in daily use at a multi-billion-dollar manufacturer
Questions We Hear First
Straight answers, before you ever get on a call.
What kinds of systems do you build?
Production software with AI inside: retrieval (RAG) systems over your documents and data, multi-agent platforms that run workflows end to end, MCP integrations into ERP and CRM systems, full-stack web applications, and applied machine-learning models. If it needs to survive daily use by a real team, it's our kind of project.
How do you keep AI agents from acting without oversight?
It's designed in, not bolted on: every consequential action — a price, a quote, an order, an outbound message — sits behind a human-approval gate. Retrieval systems cite their sources and refuse to answer when confidence is low. Decision logs record what the system proposed and what a person approved.
Can you integrate with our ERP and CRM?
Yes — that's most of the work we do. Current and recent integrations include Microsoft Business Central, HubSpot, GoHighLevel, Outlook/Microsoft Graph, calendars and telephony, often via Model Context Protocol (MCP) servers with tightly scoped access. Systems run in your cloud tenant, on your accounts.
Who owns the code, the infrastructure and the data?
You do. We build in your repositories and your cloud environment, intellectual property transfers as the work is paid for, and handover with documentation is part of every project — no lock-in by design.
Why are your case studies anonymized?
Because our contracts promise clients confidentiality, and we keep that promise — it's the same discipline we'd apply to your data. Named references are available on request once we're in a real conversation.
How does an engagement start?
You book a short call and tell us what you're trying to build. It's a technical conversation from the first minute — with the engineer who would do the work. If it makes sense, you get a written scope with acceptance tests per deliverable, and a quote. No sales team, no handoffs.
Tell us what you're trying to build.
Book a 30-minute call and describe the problem in your own words — we'll come with questions worth asking, and you decide if it's worth building.
You'll be talking to the engineer who would build it.