Skip to content
Holits
Menu

Labs

Open source and experiments.

Independent projects where I test ideas in distributed systems. Some are open source, some run on private infrastructure. The engineering reasoning is documented either way.

Send me your setup
AI InfrastructurePrivate

LLM Router

A Go routing service that controls concurrency across llama.cpp workers, choosing whether a request runs now and where — based on live load, queue depth, latency and measured token throughput.

Problem
GPUs running LLMs are not web servers, but everything routes to them as if they were. Conventional load balancers spread traffic evenly and make tail latency worse.
Solution
A stateless Go router with Redis as the sole queue and state backend, scoring every eligible worker on load, queue depth, latency, throughput and KV pressure before admitting work.
  • Go
  • Redis
  • llama.cpp
  • Prometheus
  • Grafana
  • Docker
AI InfrastructureOpen source (MIT)

Infrastructure MCP

A self-hosted Model Context Protocol server that gives AI agents safe, audited access to real infrastructure — high-level tools instead of raw SSH.

Problem
Teams want AI agents to help operate infrastructure, but the obvious implementation — give the agent shell access — has an unbounded blast radius and no audit trail.
Solution
An MCP server exposing inventory-driven, validated, structured operations such as disk_usage() and docker_restart(), where the agent reasons and the server is the only thing that touches infrastructure.
  • Go
  • MCP
  • Docker
  • Kubernetes
  • Proxmox
  • Grafana
Blockchain InfrastructurePrivate

Ethereum RPC Router

An intelligent load balancer for Ethereum RPC nodes with canonical-data guarantees, re-org detection, client-aware routing and native WebSocket proxying.

Problem
Ethereum RPC looks like ordinary HTTP but isn't: methods have wildly different costs, not every node can answer every request, and a lagging node returns answers that are quietly wrong.
Solution
A HAProxy-fronted router with a Node Weighter middleware that classifies every method, detects client type and block height, and routes to the cheapest node tier that can answer correctly.
  • TypeScript
  • Bun
  • HAProxy
  • Erigon
  • Geth
  • Reth
AI InfrastructurePrivate

Multi-GPU LLM Provisioning

An Ansible playbook that deploys one llama.cpp inference server per GPU on a multi-GPU host — each with its own model, port, parameters and systemd unit.

Problem
Standing up a multi-GPU inference host by hand is slow, undocumented and impossible to reproduce. When the box needs rebuilding, the configuration lives in someone's shell history.
Solution
A single playbook that maps each GPU to a model, renders a systemd unit per instance, and is validated by syntax-check and check-mode runs in CI before any deploy.
  • Ansible
  • llama.cpp
  • systemd
  • GitLab CI
  • NVIDIA
  • Linux
AI InfrastructurePrivate

Upgrade Intelligence

An LLM-driven dependency-update platform that turns raw changelogs into scored, bundled upgrade plans — and checks whether your codebase actually uses the APIs that are about to break.

Problem
Automated dependency bots tell you a new version exists. They do not tell you whether it matters, what breaks in your specific code, or which of the twelve packages that just bumped together actually need to move as a unit.
Solution
A Bun/TypeScript platform that fetches changelogs, scores upgrade urgency from real signals, scans a connected codebase for usage of soon-to-be-removed APIs, and bundles related upgrades into a dependency-ordered sequence — then correlates CI results back against its own predictions.
  • TypeScript
  • Bun
  • Express
  • PostgreSQL
  • Prisma
  • BullMQ
Blockchain InfrastructurePrivate

CEX Tracker

A multi-exchange monitoring service that watches nine centralized crypto exchanges for new listings and announcements, and pushes real-time alerts to Telegram, Discord and a REST API.

Problem
A new trading-pair listing is a short-lived edge, and it is scattered across nine exchanges that each publish it differently — some through a clean API, some only through a web page actively hostile to scraping.
Solution
A Node.js service that polls each exchange through the fastest reliable channel available — direct API where one exists, browser-based collection only where no API exists — normalizes the result, and fans it out through Telegram and Discord bots and a rate-limited REST API.
  • TypeScript
  • Node.js
  • MongoDB
  • Redis
  • Express
  • Puppeteer
Blockchain InfrastructurePrivate

Mempool Analyzer

A 13-service Ethereum mempool analytics platform that decodes pending transactions before they're mined, surfacing MEV, gas-market and DeFi liquidation activity in real time.

Problem
Block explorers show you what already happened. MEV, gas-market moves and liquidations are decided in the seconds a transaction spends pending — and by the time it's confirmed, the information is historical.
Solution
A message-bus-connected microservice pipeline that subscribes to a node's pending-transaction feed, decodes calldata through an ABI-then-heuristics waterfall, and fans the result out to dedicated analytics services for swaps, liquidations, bridges and gas auctions.
  • TypeScript
  • Bun
  • RabbitMQ
  • MongoDB
  • ethers.js
  • PM2
Platform EngineeringPrivate

Company Intel

An EU VAT validation and company-registry API built around the failure modes of VIES, not just its happy path — circuit-broken per country, with a legally usable audit trail on every check.

Problem
VIES, the EU's VAT validation service, is free — and unreliable enough per member state that 'call the government endpoint' is not a product on its own. Company registry data, meanwhile, is scattered across a dozen incompatible national formats.
Solution
A FastAPI service that wraps VIES with a per-country circuit breaker and stale-data fallback, normalizes company registry data from multiple national sources behind one canonical identifier, and issues an audit-proof consultation number on every VAT check.
  • Python
  • FastAPI
  • PostgreSQL
  • SQLAlchemy
  • Redis
  • RQ

Write-ups

AI infrastructure case studies

AI serving and agent tooling is explored here rather than sold as a service.

Track record

What I have built

Beyond the systems documented above, the work behind them falls into these categories.

  • Cloud Platforms
  • AI Infrastructure
  • Blockchain Infrastructure
  • Internal Developer Platforms
  • Automation Platforms
  • High Availability Systems
  • Developer Tooling
  • Observability Platforms
  • Infrastructure Control Planes