Selected Work

Eight sectors, two paradigms, in-house frameworks.

Cloud and on-prem. Three with a full case here; all eight under Industries. Not one project: a repeatable method with reusable assets.

AviationProductionData & BI platform — European airlineFrom zero to the company's single source of data, declared by the presidency.36Dashboards in production, each with its runbook · 10 departments + management136Operational sources —SFTP, APIs and databases— across 400 artifacts in three layers~€9.5Infrastructure cost per user per month ($4,767 over twelve months, 50 users)$4,700–$30,700Annual saving against conventional alternatives, depending on scenario (2–7× more expensive, official AWS prices)
FinanceProductionOn-prem financial lakehouse~740Tables in the core schema, documented for the first time in the cooperative's historyOwn IPProductioncortex-ai — one agentic routing engine, two products in production2products in production on the same engine, in unrelated domains — one verifiable live on this site

More work

Depth on demand: the rest of the work, with its honest status. Product, distribution, meta-tooling, and in-house agentic AI.

ProductMVP in progressAutonomous marketing platformA Stack Overflight product, not my own: a SaaS for autonomous travel-campaign optimization — "algorithmic trading applied to advertising". I designed the data architecture, all the infrastructure and the Go API; I'm still building it. It's a full Go core running on AWS Lambda via Function URL, with no persistent server. The decision core is cortex-ai, the agentic routing engine I designed for Luna Azul —deployable as a backend in any product— which travel-marketing adopted, not the other way around.
Meta-toolingProductionCore — Medallion ETL frameworkA Medallion factory-platform: base classes, a declarative registry, and a handler factory. Adding a fact = implementing one method. ~90 KLOC of core support 185+ data modules operated by a team and agents.
Meta-toolingProductiontf-admin — agent-first infra CLIA unified Terraform admin CLI: 1 entrypoint, ~40 commands, multi-project. Designed agent-first so an AI agent can operate infrastructure safely, with human guardrails on destructive actions. ~22 KLOC, solo.
Meta-toolingProductionbi-ai-tool — BI automationI wrote the client that operates Superset and Grafana via API: a shared generic REST core —two backends, one base client, JWT+CSRF for one, Bearer for the other— with per-project multi-tenant isolation. Every change is written down, reviewable and reversible. It's operated by the Armando agent.
Meta-toolingProductionazure-devops — IaC for project managementThe same Terraform pattern —declarative spec, diff detects drift, apply reconciles— applied to project management itself: sprints, teams, work items, wiki, dashboards in Azure DevOps. Infrastructure as code applied to management: the desired state lives in versioned files and the board is the result, with the same verbs (import, diff, apply --dry-run) I use for real infrastructure. It's operated by a human or the Yuki agent.
Own IPProductionIn-house ecosystem — Luna Azul + multi-tenantA cluster of legal and administrative agents in real production (they review contracts, not demos), with a meta-agent that designs them and a multi-tenant platform with persistent, multi-channel memory.
DistributionDeliveredDepartmental data lakeA greenfield on-prem build where no data infrastructure existed: reverse-engineering the ERP, config-driven ETL, sales-forecasting ML with Prophet, and 13 business domains. Full-stack data work, solo.
DistributionProductionSQL Server optimization — Siesa ERPA bottleneck that looked like hardware and wasn't. A distributor's server on the Siesa ERP had 64 vCPUs and uncapped dynamic memory, yet the real workload used just one of its cores. I diagnosed it with my agents Atlas, Charys and Veldora from the physical layer down to the query optimizer: static memory and 16 vCPUs, 128 fragmented indexes rebuilt to 0 in 22 minutes, and backups redesigned from 8 full daily copies into a tiered scheme. Intuition blamed the query engine; an A/B on an isolated copy refuted it —the optimizer underestimated cardinality by two orders of magnitude and chose a serial plan no matter how many cores were available—. Migrating to NVMe gives 25x on sequential operations but only 1.4x on the latency that actually hurts billing: it helps, it's not the silver bullet. Diagnosis and fix: ~3 hours.