AI tutoring startup
Senior Software Engineer
Replaced message-count entitlements with cost-based usage metering and built the AI tutor's memory pipeline, on a platform used by millions of secondary-school students.
Borut Balazek
Senior Software Engineer
Backend & AI Engineering
I've built and scaled backend systems for 15+ years, eight of them on ClassCharts, where I started as the first developer and stayed until it served around 1,000 requests a second. Now I build AI features and what they run on: usage metering, queues, retrieval and data pipelines.
- TypeScript
- Node.js
- PostgreSQL
- Python
- Redis
- Data pipelines
- LLM systems
- Kubernetes
- AWS and GCP
- All skills →
Remote · Slovenia · English and German
Open to a long-term engineering role or freelance backend and AI projects.
Experience
If your system has to keep running while it grows, these are the teams where I did that, and what I built for them.
MarketAlerts
Founding Engineer
Took over an agency-built investment-intelligence platform and worked across its web and mobile apps, AI layer and infrastructure until the company wound down in March 2026.
- 20+ global markets
- Web, iOS, Android and the API
Smart-city IoT platform
Senior Backend Engineer
Designed the backend of a smart-city programme: 20+ microservices on Kubernetes across five city domains, built in parallel by several vendor teams.
- 20+ microservices
- Five city domains
Edukey Education / TES Global
Senior Software Engineer
Joined ClassCharts as the first developer alongside the CTO and stayed eight years as it grew past ten million registered pupil, parent and teacher accounts, through its acquisition by TES Global.
- ~1,000 requests a second
- 40+ TB migrated between clouds
- Six store listings, ~2M downloads
Corco Labs
Independent Software Engineer
My own company. Around 70 client projects since 2012, many of them through agencies for brands like Philips, RWE and Red Bull.
Visit Corco Labs : Corco Labs →What people I've worked with say
Colleagues, managers and founders I've worked alongside, several of them over many years.
" I worked with Borut for 8 years, he was our longest serving developer and he was an excellent person to work with. Borut always feels & displays a high level of responsibility to the project and the team and brought immense value to what we were trying to achieve, he understood our vision & grew with our codebase to become a hugely valued developer who you could rely on & would always deliver. Time keeping, attention to detail & working with others were never a problem. Borut worked on each of our products, working with different team members, platforms and codebase & responsibilities. I have zero hesitation in recommending Borut as he is an intelligent & hard working individual with a good sense of humour and liked by the rest of the team - he was a huge asset! " " Borut delivered immediate value after joining our team. He brought clarity to complex problems, unlocked efficiency, and significantly increased throughput. He quickly identified mission-critical technical debt and resolved it on schedule. Borut communicates clearly with both technical and non-technical stakeholders and consistently connects technical concepts to practical business goals. His experience enables him to leverage AI to accelerate delivery without compromising quality. We were fortunate to have him on board and would gladly collaborate with him again or recommend him for any project. " AI work
If your team is putting language models into a product, these are the parts I have already built: what each call costs, what the model remembers, what it retrieves and what an agent is allowed to change.
- AI tutoring platform · client
Usage metering by token cost
Each request is priced by the tokens it uses, checked before the model call and charged only once the call succeeds, across every plan tier.
How it works : Usage metering by token cost → - AI tutoring platform · client
Long-term memory for an AI tutor
Lasting facts are extracted from conversations in the background, then merged and pruned every night, with the load spread so the database stays flat.
How it works : Long-term memory for an AI tutor → - Metis · my own platform
Retrieval over news at scale
Around 100,000 articles a day are embedded into Qdrant, searchable by meaning and available to an LLM agent as tools before it makes a call.
How it works : Retrieval over news at scale → - MarketAlerts · founding engineer
Classification with a cheap first pass
A cheap model takes the first cut and a stronger one runs only where the call is hard, behind one registry of providers that can be swapped.
How it works : Classification with a cheap first pass → - Guestavo · my own product
An MCP server with guardrails
An operator runs a venue from an AI client, while every change an agent makes is logged with its before and after state and can be undone.
How it works : An MCP server with guardrails → - My own workflow
Building with coding agents
I work with coding agents every day, with every change verified and reviewed by a second model and by me. The workflow is public as agent skills, and in my own projects an orchestrator runs standing agents in isolated git worktrees.
How it works : Building with coding agents →
Controls around the model calls
-
Cost checked before the call
Requests are metered by what they cost in tokens, checked before the model runs and charged only on success.
-
Agent actions can be undone
Changes made over MCP are logged with their before and after state.
-
Providers can be swapped
Models sit behind a registry that keeps each provider's config intact, and one summariser moved from OpenAI to Gemini in production.
Need help putting AI into your product or team workflow? I offer scoped consulting pilots through Corco Labs.
Explore AI consultingHow I work
What happens to a change before it reaches your users.
- I read the existing codebase before proposing anything, and build in its patterns instead of adding a second way to do the same thing.
- A fix is done when I have run the real path and checked the output: the API response, the queue depth, the console errors and the admin tables.
- Each pull request gets an automated review and my own, and I work its comment list down to zero.
- Production changes go out in small steps: a batch job runs by hand against one real user before it runs for everyone, and new usage limits are logged in shadow before they block anyone.
- Decisions are written down where the next person will meet them, as a comment above the line or a dated decision record, and a mistake that keeps coming back becomes a lint rule.
My projects
If you want to see what I build when the whole product is mine, these are products I designed and built myself.
Guestavo
2026A guest-engagement platform for independent restaurants, bars, cafes and event-led venues: reservations, digital menus, events, campaigns, loyalty, staff scheduling and a guest portal that closes the loop after a visit. A Next.js operator app and public venue pages, a Fastify API, a BullMQ worker, PostgreSQL via Drizzle and an MCP server for AI clients. Live in beta; designed and built by me in 2026.
TypeScript Next.js Fastify PostgreSQL Redis
Website : Guestavo →
AfterGather
2026A media hub for photographers that joins two workflows usually split across separate tools: private client galleries with proofing, and QR-based guest photo and video uploads at live events. A Next.js app, a Fastify API with TypeBox schemas, a BullMQ worker for virus scanning and video transcodes, PostgreSQL via Drizzle and S3-compatible storage, with billing on a credit ledger. Designed and built by me in 2026, with access through a public waitlist.
TypeScript Next.js Fastify PostgreSQL BullMQ
Website : AfterGather →
Metis
2021 - presentA personal platform that takes in around 100,000 financial news articles a day from 5,000+ feeds, newsletters and archives and lines them up against real-time market prices. An event-driven pipeline on BullMQ and RabbitMQ moves each article through self-hosted models for embeddings, clustering, entity recognition and sentiment, and into storage split by job: PostgreSQL, ClickHouse, Meilisearch, Qdrant and Redis.
TypeScript NestJS ClickHouse Meilisearch Redis
Read the case study : Metis →
Pocket AI Gateway
2026A self-hosted gateway for OpenAI, Anthropic and Gemini clients. One Go process serves the API and an admin dashboard and keeps configuration and usage in SQLite. Stable model names route to providers by fixed, fallback, weighted, cost or latency strategy, with scoped API keys, request, token and spend limits, and cost traced per request. Open source under MIT, in alpha.
Go TypeScript Next.js SQLite Docker
GitHub : Pocket AI Gateway →
Two more products are in stealth, one in EdTech and one in health.
Writing
Notes on what I build and how I work, here and on Substack.
-
🏫 Eight Years on ClassCharts
A few engineering decisions from my eight years on ClassCharts.
7 min read
-
🗂️ My Agent Skill Tree
A tour of my coding agent skills and how they fit together.
10 min read
-
📰 Metis Revisited: The Platform Behind the News
An update on Metis and its financial data tools.
8 min read
-
🧠 Five Attempts at Giving an Assistant a Memory
My five attempts at building a personal assistant that remembers.
9 min read
A bit about me
The journey so far
I started coding back in 2004 with simple HTML sites, made it my job in 2011, and have since built systems that handle millions of users. Along the way, I've discovered that the best software isn't just about clever code - it's about solving real problems for real people. What teams tell me they value is that I take ownership, communicate honestly about tradeoffs, and stay with a system long enough to be accountable for it.
These days, I work mostly with TypeScript, React, Node.js, PostgreSQL, Python, Docker and Kubernetes, across product code, data pipelines and infrastructure. I use modern AI-assisted development workflows where they improve delivery without compromising engineering quality, and I've also built LLM-backed product features - usage metering, memory, classification pipelines, alerting and search - as part of real systems.
From my own projects like Guestavo, Metis and MoaiTime, and WeWentCarRacing (because I love motorsports), to work with Edukey and TES Global, through agencies for Philips, RWE, Innogy, Red Bull, Danone, Les Mills and Radenska, and a smart-city platform - I enjoy building things!
What drives me
- Building things people love - There's nothing quite like seeing someone actually use and enjoy something you've built. That's what keeps me coding.
- Learning constantly - The tech world moves fast, and I love keeping up. Whether it's a new framework or an old problem with a fresh approach.
- Sharing knowledge - I write about my projects, contribute to open source, and enjoy helping other developers when I can.
- Ownership over tickets - I'd rather be accountable for a system end to end than hand work over the wall and hope it lands.
- Making things simple - The best software is invisible to the user. Complex problems shouldn't require complex interfaces.
- Integrity in communication - I say what I know, what I don't, and what I'm doing next, so the people I work with always have the full picture.
I've been remote since the start of my career in 2011, long before it was normal. Fifteen years of clients and teams in the UK, Germany, Austria, Switzerland and the US, across time zones and mostly in writing, means a distributed team is the setup I know best.
Outside work
- Music production and DJing
I started producing and DJing electronic music as a teenager and came back to it recently. Some productions are on music.bobalazek.com, and I wrote about the journey and my library.
- 3D printing
My 3D printers turn ideas from my other hobbies into real parts. The biggest build so far is a DIY formula-style sim racing steering wheel, which I wrote up prototypes and all.
- Motorsport and sim racing
I follow racing closely enough to have built WeWentCarRacing, a calendar across 20+ series. I race on the sim too, which led to WeWentSimRacing and a post about my setup.
- Open source
I wrote the physics helper in the Babylon.js 3D engine, and it still ships. I also publish my own projects, like MoaiTime, OpenVoice Flow and my agent skills.
🇩🇪 Übrigens, ja, ich spreche natürlich auch Deutsch! Diese Seite gibt es auch auf Deutsch.
Work with me
For a long-term engineering role or a freelance backend or AI project, send the role or the problem, your stack and your timeline. I'll tell you whether I'm a good fit.
"It has been an absolute pleasure working with you over the past five years. In 25 years of coding, you are one of the most gifted developers I have ever worked with and an all round great guy." Borut Balazek
Senior Software Engineer, Backend & AI Engineering
- Ljutomer, Slovenia (CET)
- English, German and Slovenian
- Open to roles and projects