A small studiobuilt around senior people.
Serge, who founded the studio, has spent 15 years putting software into production, at places like Ember and MortgageBloc. Repetitive code goes to AI and the decisions stay with senior engineers, so the team can stay small and still move quickly. One person answers for how your project turns out, and you'll know his name from the first call.
- Approach
- Problem first, then the stack
- Tools
- Claude Code, custom agents and MCPs, evals
- Result
- Production software in weeks, at a fixed price
How we work
The steps don't change much from one project to the next, and there are four of them.
Start with the problem
Nobody opens an editor until we've pinned down the outcome you need and what's in the way. Now and then a brief asks for more than the problem needs, and we'll tell you when a smaller fix would do.
Design the system
Then we sketch the architecture, the data model and the user flows around that outcome. The foundations use proven tools, so your next hire won't have to learn anything exotic.
Build with AI, review by hand
Claude Code and our own agents write the scaffolding, tests and boilerplate. A senior engineer reads every change and owns the result. We call the split 90/10.
Ship, then stay
Within weeks, working software is running in your own repo and infrastructure. We stay for 30 days after launch, which is when your first users find the things that break.
AI-native,human-led.
AI lets a small team like ours get through a lot more work. Where it goes in your product, and where it stays out, is a call a senior engineer makes.
- Claude Code and agents
- We've used it every day since 2025, currently with Opus 5 and Fable 5.1. We also build agents for your product, with approval steps and audit logs so you can see what they did.
- Custom MCP servers
- We write Model Context Protocol servers that connect AI to your codebase, database and internal systems, with permissions a security reviewer would sign off on.
- Evals and observability
- We ship each AI feature with a test set and monitoring, so you see a drop in quality before a customer emails you about it.
- Cost control
- We route each call to the smallest model that does the job and cache what repeats, so your AI bill stays a predictable line.
What we build
It's TypeScript from the database to the browser, written so your next hire can read it on their first day.
- AI agents and copilots built into working products
- Full-stack web applications (Next.js, TypeScript)
- Mobile apps (React Native)
- Internal tools to replace the shared spreadsheet and the endless email thread
- Connections to payments, CRMs, ERPs, WhatsApp and Telegram
- Rescues for MVPs built with AI that now have to cope with real users
Who you
talk to.
You'll speak to the founder on the first call, and he stays on your project until it's done. If part of the build needs a specialist, he's the one who brings them in.
Serge Radkevics, founder
Before SoulToSoul, Serge was Tech Lead at Ember, a UK startup automating accounting that Starling Bank went on to acquire. He was also Tech Lead at TribeIRL, and CTO of MortgageBloc. That adds up to fifteen years in fintech, logistics and consumer apps. He works from Portugal.
- Experience
- 15+ years shipping production software
- Stack
- TypeScript, Next.js, NestJS, React Native, Claude Code
- Roles
- CTO, Tech Lead and Architect in fintech, logistics and consumer apps
What we work withas of September 2026
We run each item on this list in production for a product we built. We review the list every month and drop what we no longer use.
AI agents and coding tools
These tools are where our speed comes from. Senior engineers tell them what to do, read what they produce and answer for the result.
- Claude Code with Claude Opus 5 and Fable 5.1 as the daily build tool
- Long-running agents with tool use, human approval steps and audit logs
- MCP servers that connect models to your codebase, database and internal systems
- Evals and monitoring, so you see an agent getting worse before a customer tells you
- RAG and knowledge systems over your own documents and data
- Voice agents for support and intake (ElevenLabs, Twilio)
AI cost and control
An AI bill creeps up when nobody is watching which model each call goes to. We find out where your money is going, and then we cut it.
- Model routing: frontier models only where the task needs them
- Prompt caching, batching and small or on-device models for the routine calls
- Spend reporting broken down by feature
- Guardrails against prompt injection and over-permissioned agents
EU AI Act and security
The transparency rules have applied since 2 August 2026. The Annex III high-risk rules follow on 2 December 2027. We can tell you which of them cover your product.
- Article 50 transparency: AI disclosure and labelling of AI-generated content
- Machine-readable marking of AI content before the 2 December 2026 cut-off for existing systems
- Risk classification under Annex III and a documented compliance file
- Security review of AI-written code: auth, data access, secrets and dependencies
- Code audits before a raise or acquisition, written for technical due diligence
Product stack
Everything is typed end to end. For the foundations we stick to proven tools, so the AI features sit on a base your next hire already knows.
- TypeScript, Next.js 16, React 19, React Native
- NestJS and Postgres (Neon), Prisma
- Vercel, Fly.io and AWS deployments with CI/CD from day one
- Stripe, Slack, Telegram, WhatsApp and email integrations
- Analytics, error tracking and dashboards that a founder can read
Tell us
the problem.
After 20 minutes you leave with a plan and a price range. Then you decide whether to build it with us.