AI systems · Custom software · Web3 · Game backends

Yarvixo builds AI software that holds under production load

The same engineering that runs 52,000 concurrent players at 62 ms average round-trip and 99.98% uptime — now applied to AI products, custom software and the Web3 work around them. Senior engineers only, and you own the IP from day one. Two of the cases below are named and public — go and check them.

87%
Less manual processing

An enterprise AI platform across 14 sources and 5 languages, for one anonymised client.

99.2%
Extraction accuracy

Same system. A measurement, not an estimate — an evaluation harness was built to take it.

52 000
Concurrent players · peak

A different system: one anonymised game backend, in production, at 62 ms average round-trip.

63/100
Metacritic · Switch

King's Bounty II. Graded by reviewers, not by us.

Trusted by companies across Europe & North America

01

Work

The first three are named, so you can go and look at them — a retail console game you can buy in the eShop, a client game live on Google Play, and our own product. The other three are client work and stay anonymised until we have written permission.

A town square in King's Bounty II: guards, market stalls and a lit portal under heavy foliage

Screenshot: Nintendo eShop listing. King’s Bounty II — 1C Entertainment / Prime Matter.

Console port — named client

King's Bounty II — Nintendo Switch Port

An Unreal Engine 4 open world built for PC wanted about 7 GB of memory and ran at 6–8 fps on a devkit. A Switch has 4 GB on the whole device. Nothing in the game was allowed to be cut to close the gap.

Five engineers, twelve months: texture pool budgets, LOD and asset repacking to fit memory, a character streamer rewritten from scratch once the CPU became the limit, and a fixed resolution set separately for docked and handheld — pixels spent so the content did not have to be.

Result: shipped day and date with PC, PS4 and Xbox One on 24 Aug 2021 · 30 fps target from 6–8 fps, with nothing cut · still on sale, and supported on Switch 2

Metacritic by platform — one game, four ports, four teams
Xbox One73
Switch63
PC63
PS457

The Switch column is ours. It lands level with the PC build and six above PS4, on hardware with a fraction of the memory of either.

Read full case study →
Three Metarun heroes racing down a lamplit street, each ringed by a speed trail

Screenshot: Google Play listing. Metarun — Encore Games.

Game & Web3 — named client

Metarun — 1v1v1 Web3 Mobile Runner

A runner is single-player by default. Making it a ranked 1v1v1 race over mobile networks means the device can no longer be trusted with the result, and in a title with tradable assets an exploited race is something the exploiter keeps.

Six engineers, four months of active development: Unreal Engine client for Android and iOS, server-authoritative match services deciding the race off-device, MMR matchmaking and ranked leagues, anti-cheat, and hero ownership on BNB Chain — ownership on-chain, gameplay off it.

Result: 39,000 concurrent players at its event peak, 60 ms average round-trip, 99.8% uptime in production · live on Google Play with 50,000+ downloads and 4.0 from ~2,970 reviews, per Google · named with the client's permission, so you can install it and check

Read full case study →
The ArendaOS dashboard: cars on rent, fleet utilisation, customer debt, and today's returns, overdue payments and service due

Our own product. Dashboard rendered from the live ArendaOS markup, in English — illustrative rows, no real customer data.

SaaS — our own product

ArendaOS — Fleet Rental SaaS

Car-rental operators run deposits, debts, fines, damage disputes and maintenance out of spreadsheets and group chats, and nobody can answer who owes what today.

A multi-tenant B2B subscription platform on NestJS and PostgreSQL — twelve modules, seven roles, PDF contracts, encrypted personal data, and overlapping bookings refused by a database constraint.

Result: shipped and live at arendaos.com · 12 modules full-stack and multi-tenant · zero double bookings, enforced in the database

Read full case study →

Anonymised client work — no name, no link to verify, listed as such

GameDev

Multiplayer Game Backend

Go microservices, Redis pub/sub, Kubernetes autoscaling, custom matchmaking. 52k peak CCU · 62 ms · 99.98% uptime.

 Read → Web3

Decentralised NFT Marketplace

Gas-optimised contract suite on Polygon with subgraph indexing and real-time feeds. $2.4M first-month volume, per the client.

 Read → AI SaaS

Enterprise AI Analytics Platform

Multi-model LLM pipeline with custom RAG across 14 data sources in 5 languages, wired into an existing ERP. 87% less manual processing, 99.2% extraction accuracy, live in 11 weeks.

 Read →

Browse all case studies →

02

Where the numbers come from

52,000 concurrent players, 62 ms average round-trip and 99.98% uptime are not three wins from three projects — they are one backend, and this is that backend. Each figure carries the condition the case study records it under.

Drag the drawing sideways →

Architecture of the 52,000-player game backend Players on Unity and Unreal clients connect through an edge load balancer to matchmaking and an autoscaled fleet of authoritative game servers, which keep presence and room state in Redis and accounts and progression in PostgreSQL. The backend reached 52,000 peak concurrent players, 62 milliseconds average latency under load, and 99.98 percent uptime after rollout. CLIENT EDGE SERVICES STATE 52 000 CONCURRENT · PEAK in production, on this system Players Unity and Unreal clients mobile, PC, console 62 ms AVERAGE under load Edge + TLS Load balancer connection draining 99.98% UPTIME after rollout Matchmaking MMR, ranked leagues, anti-cheat Game server fleet authoritative, Go Kubernetes autoscaling Redis pub/sub, presence, rooms PostgreSQL accounts, progression Every box and every condition above is stated on the case study or in the architecture article. Nothing here is inferred.
The 52,000-player deployment: Go services behind an edge tier, matchmaking, an autoscaled fleet, Redis and PostgreSQL. Autoscaling runs on connection count rather than CPU, and the live ops team was handed Grafana dashboards before launch. The long version is here.
03

What we build

AI and custom product work first, with the real-time and Web3 engineering that evidences it.

01

AI development

LLM integration and fine-tuning, RAG systems, agents and pipelines — built into a product that already exists, rather than bolted alongside it. The platform behind the 87% above is a custom RAG across 14 sources in 5 languages.

PyTorchRAGLLM fine-tuningAgents AI development → 02

SaaS development

Multi-tenant B2B platforms, API-first, MVP in 8–12 weeks. ArendaOS is ours end to end and is the one you can open in a browser today.

NestJSPostgreSQLReact / Next.jsAWS · GCP SaaS development → 03

Smart contracts

Contract engineering, dApps, NFT infrastructure and game economies. On Metarun the split was deliberate: ownership on BNB Chain, gameplay off it.

SolidityBNB ChainPolygonSubgraphs Smart contracts → 04

Game development

Multiplayer systems, game backends and full-cycle production in Unity and Unreal Engine — including console ports through certification. King's Bounty II shipped to Switch on this work.

UnityUnreal Engine 4 / 5Switch · PS4 · XboxNetcode Game development → 05

Multiplayer backends

Authoritative servers, matchmaking, ranked leagues and anti-cheat. The 52,000-CCU deployment above runs on Go microservices with Redis pub/sub and Kubernetes autoscaling.

GoRedisKubernetesPostgreSQL Multiplayer backends →
04

The stack

What we actually reach for, grouped by where it runs. We are not dogmatic about tooling — but these are the things we have shipped and can support at three in the morning.

Game clients

  • Unreal Engine 4 / 5
  • Unity
  • C++
  • C#
  • Nintendo Switch
  • PlayStation
  • Xbox
  • iOS / Android

Servers & real-time

  • Go
  • Node.js
  • NestJS
  • Python
  • WebSockets
  • gRPC
  • Authoritative netcode
  • Matchmaking / MMR

Data

  • PostgreSQL
  • Redis
  • MongoDB
  • ClickHouse
  • S3
  • Exclusion constraints

Infrastructure

  • Kubernetes
  • Docker
  • Terraform
  • AWS
  • GCP
  • CI/CD
  • Prometheus
  • Grafana

AI

  • PyTorch
  • RAG pipelines
  • LLM fine-tuning
  • Vector search
  • Agent orchestration
  • Evaluation harnesses

Web3

  • Solidity
  • BNB Chain
  • Polygon
  • Ethereum
  • Subgraphs
  • Contract audits
05

How we work

Six stages, transparent at each one: weekly demos, live dashboards and direct access to the engineers doing the work.

01

Discovery

Requirements, scoping, team setup. A written scope with a fixed price against it within 48 hours of the call.

02

Architecture

System design and the technology decisions, made before anyone writes production code.

03

Development

Agile sprints with mandatory code review. Weekly demos of something that runs.

04

Testing

QA, security and performance audits — unit, integration and end-to-end, in CI.

05

Launch

CI/CD and production deployment, with the runbook handed over alongside the code.

06

Support

Monitoring, incident response, scaling and iteration — or a full handover to your own team.

Let's build your next product

Two of our cases are named and public: you can buy one in the eShop and install the other from Google Play. Start with a call and we will tell you what we would actually do — and what we would not.

06

Writing

The questions that come up on almost every first call, answered in public at the length they actually need. No gated PDFs.

Architecture

Game backend architecture for 50,000 concurrent players

What the stack looks like, why each part is there, and the four things that break first on the way up.

 Cost

What does a multiplayer game backend cost to build and run?

About $1,200 a month at 500 peak CCU, $6,900 at 5,000 and $48,000 at 50,000 — line by line, on list prices.

 Engines

Unity or Unreal Engine 5 for a multiplayer game?

Unreal 5 ships production replication and is the default for high-fidelity PC and console. Unity wins elsewhere, and where is the point.

 AI

RAG or fine-tuning: which does your product need?

Almost always retrieval. Fine-tuning changes how a model behaves; retrieval changes what it knows. One question tells you which you have.

 Web3

Web3 in games: what actually belongs on-chain

Ownership belongs on-chain. Gameplay does not. Where the line sits, and what putting the game loop on a chain costs you.

 Databases

Preventing double bookings in PostgreSQL

An exclusion constraint instead of application validation — the approach behind the zero double bookings in ArendaOS.

All eleven articles →

07

Questions

Everything worth knowing before we start working together.

An AI SaaS MVP runs $40,000–$120,000, a SaaS MVP $40,000–$90,000, and an LLM integration into an existing product $15,000–$30,000. Multiplayer backends and Web3 marketplaces are scoped per project rather than banded, because peak concurrency and audit scope drive them. Every band we publish, the three engagement models and what moves the number are on our pricing page. After a free discovery call you get a written scope with a fixed price against it within 48 hours.
Most MVPs are delivered in 8–12 weeks depending on complexity. We use a lean sprint model — you get a working, deployable product with core features, not a prototype. We can start within 1–2 weeks of signing.
Absolutely. About 60% of our clients are early-stage startups. We understand the need to move fast, stay lean, and build something investors can evaluate. We also have experience working with equity-based or deferred payment models for seed-stage companies.
We offer three models: Managed Project Delivery (outsourcing — we own the delivery), Dedicated Team Extension (outstaffing — your engineers, our talent), and Tech Talent Acquisition (recruitment into your own team). Most clients start with outsourcing and transition to outstaffing as they scale.
You do — 100%, from day one. We sign IP assignment agreements as part of our standard contract. There is no vendor lock-in, no licensing fees, and no restrictions on how you use, modify, or sell the product we build for you.
Our teams operate in CET and US Eastern time zones, giving us real daily overlap with clients across Europe and North America. We maintain async-first communication with synchronous standups and weekly demos scheduled at your convenience.
Yes. GDPR compliance is built into our delivery processes by default — data handling, storage architecture, consent flows, and DPA agreements are all standard parts of how we work with EU and UK clients.
A typical AI-powered SaaS MVP ranges from $40,000 to $120,000 depending on the complexity of the AI pipeline, number of integrations, and required infrastructure. We provide a detailed breakdown after a free discovery call — usually within 48 hours.
Outsourcing means Yarvixo owns the delivery — we provide a full team, manage the process, and deliver the product. Outstaffing means you hire individual engineers from our talent pool who work as an extension of your own team under your management. Both models are available and can be combined.
A basic smart contract with a frontend dApp takes 4–8 weeks. A full NFT marketplace or DeFi protocol with audited contracts, multi-chain support, and a polished UI typically takes 12–20 weeks. Timeline depends heavily on audit requirements and the number of supported chains.
You get a pre-vetted team of 2–8 engineers, a tech lead, and optionally a QA engineer — fully integrated into your workflow, tools, and rituals. The team operates on your schedule, uses your project management tools, and reports directly to you. Minimum engagement is 3 months.
Every project includes mandatory code reviews, automated testing (unit, integration, and E2E), CI/CD pipelines, and architecture review checkpoints. We follow OWASP security guidelines, enforce linting and type safety, and deliver full documentation alongside the codebase.
Yes. We sign an NDA at the start of any substantive pre-sales conversation where confidential information is shared. This is standard practice and requires no negotiation on your end — we use a straightforward mutual NDA that can be signed within 24 hours.
Our strongest domain expertise is in SaaS, AI startups, Web3 / DeFi, gaming and full-cycle game development, fintech, and HealthTech. We also have experience in logistics automation and e-commerce infrastructure. We choose projects where we can bring genuine technical depth, not just execution capacity.
Yes. Most of our client relationships extend well beyond the initial MVP. We offer post-launch support packages covering monitoring, incident response, feature development, performance optimisation, and scaling. We can also transition the codebase to your in-house team with full handover documentation.
Our default SaaS stack is React or Next.js on the frontend, Node.js or Python on the backend, PostgreSQL or MongoDB for data storage, and AWS or GCP for infrastructure. We use Docker and Kubernetes for containerisation and Terraform for infrastructure as code. We adapt the stack to your requirements — we are not dogmatic about tooling.
The fastest path is to book a free 30-minute discovery call via our Calendly link or fill in the contact form below. We will review your project brief within one business day and come prepared with initial thoughts, relevant case studies, and a proposed next step — no commitment required.
Absolutely. Many of our clients are business founders or domain experts without a technical background. In these cases we take on a stronger advisory role — helping define the architecture, select the right tools, and translate business requirements into technical specifications. You do not need to know how to code to work with us effectively.
08

Get in touch

Fill in the form or reach out directly — we respond within one business day.

Start a conversation

Whether you have a detailed brief or just an idea, we're happy to help you figure out the best path forward. No sales pressure — just an honest conversation.

[email protected]
calendly.com/yarvixo
Response within 24 hours

Message received

We'll review your enquiry and respond within one business day.

GDPR-compliant delivery
NDA on day one
EU & US timezone coverage
Full IP ownership transfer
Senior engineers only

Legal

Last updated: March 15, 2025