Case Studies

What actually happened, including the mistakes.

Written the way we would want to read them: the real failure mode, the design tension, and what we would do differently.

Book a Discovery Call

Blockchain InfrastructureWeb3 infrastructure

Architecting a 13-service pipeline to decode the Ethereum mempool in real time

Building a message-bus-connected microservice platform that turns raw pending-transaction calldata into structured MEV, gas-market and liquidation signal before a block ever confirms.

13 independently deployable services communicate through one shared message-bus abstraction, with no direct service-to-service calls
A decode waterfall (ABI, then Sourcify, then heuristics) always publishes a result, so no pending transaction is silently dropped
Enrichment — the expensive full-calldata and trace-data path — runs only on the fraction of transactions that need it
An MCP-based triage loop lets the decoder close its own coverage gaps under human-gated dry-run and auto-apply flags

AI InfrastructureAI infrastructure

Queue-based infrastructure for scoring, bundling and verifying dependency upgrades

An async worker architecture that turns changelog noise into a dependency-ordered upgrade plan, with a multi-provider LLM layer resilient to any single provider's outage.

Heavy analysis work runs on Redis-backed BullMQ queues across seven workers, keeping the API responsive under load
Related packages upgrade as one dependency-ordered bundle, sequenced with a topological sort, instead of racing as independent PRs
A multi-provider LLM adapter with a circuit breaker keeps analysis running through a single provider's outage, including local and self-hosted models
CI results are correlated back against predicted breaking changes into a single confidence score, rather than a binary pass/fail gate

Blockchain InfrastructureCrypto markets infrastructure

A tiered polling architecture for tracking nine exchanges in real time

Designing a multi-exchange monitoring platform that matches its infrastructure cost to what each exchange's API actually allows, instead of one uniform, worst-case polling strategy.

Six of nine exchanges poll through direct APIs on a 15-second cadence; the three that block direct access run on a separate, slower, cached path instead of degrading the rest
Aggregation now scales with pairs that actually changed, not total pairs tracked — cycle time down from ~50s to 5-8s as the pair count grew past 13,000
Notification delivery streams subscribers via cursor instead of loading the full recipient set into memory per send
Telegram and Discord bots share one rate limiter and one subscription-tier check, so gating stays consistent across channels

Platform EngineeringCompliance / fintech infrastructure

Building a resilient compliance API on top of an unreliable government service

A per-country circuit breaker, canonical-identifier data model, and an isolated billing service — the platform engineering that turns a free, unreliable VIES endpoint into a dependable product.

A per-country circuit breaker serves the last known-good VAT result during a VIES outage instead of an error, labelled as cached so callers can distinguish it from a live check
Every VAT check returns VIES's own consultation number, making the result usable as compliance proof, not just an internal boolean
Company records from multiple national registries resolve to one canonical identifier, derived from country and registration number
Billing runs as an isolated internal service with its own database and Stripe integration, so a billing-provider outage cannot take down VAT validation

Blockchain InfrastructureWeb3 infrastructure

Making Ethereum RPC correct under load and re-orgs

Replacing a round-robin balancer in front of a mixed Erigon/Geth/Reth fleet with a router that classifies every method and guarantees canonical data.

Light RPC methods no longer consume archive-node capacity
Erigon-only methods can no longer reach a Geth or Reth node
Responses from abandoned forks are detected instead of served as canonical
Every routing decision is labelled by method and observable in Grafana

AI InfrastructureAI infrastructure

Making multi-GPU inference predictable instead of fast

Replacing even traffic distribution across llama.cpp workers with admission control based on live load, queue depth and measured throughput.

Scheduling decisions carry a full scoring breakdown and are individually explainable
Throughput and latency averages survive worker restarts instead of resetting
Host rebuilds became a playbook run rather than a manual afternoon
Critical-priority work is no longer excluded by normal-priority queue caps

Platform EngineeringPlatform engineering

Giving AI agents infrastructure access without giving them a shell

Replacing agent SSH access with an MCP server exposing inventory-driven, validated, audited operations.

Every agent action produces a structured audit record
Mutating operations require explicit confirmation; read-only is the default
Credentials never reach the agent's context
Agents can only address hosts that exist in inventory