Portfolio

Real systems. Real constraints. Real numbers.

Six engagements across fintech, gaming, marketplaces and public-sector data — streaming and AI systems in production, live for years. The evidence behind the path: diagnose, prove, ship — and the judgment to direct it. Plus two recent solo builds to show current capability.

100K+ events/min under live ops
97.5% fraud detection accuracy
79× latency improvement, same hardware
€10M+/day in derivatives processed

Featured client work

Production systems, years live.

ESL tournament automation pipeline schema

97%time cut

97%accuracy

$0.30per process

Problem. New streaming pipelines needed bootstrapping data extracted manually from messy documents. Slow, error-prone, blocking tournament launches.

What I built. A 4-microservice architecture combining LLM-based parsing with graph-solver reasoning. Templates matched, edge cases handled automatically. Solo design and implementation.

Outcome. 2–3 hours per config dropped to 5 minutes. Errors caught pre-production. Launches no longer blocked.

Stack: Python · LLM APIs · graph solver · Scala microservices · AWS

Société Générale CIB termsheet generator schema

50%drafting time cut

€10M+/dayvolume

Zerocompliance surprises

Problem. Trading desks drafted complex derivative contracts manually — slow, error-prone, compliance risk. Each clause depended on dozens of business rules scattered across people’s heads.

What I built. A solver encoding those rules, generating contracts automatically. Lawyers review, not write. Controllers approve via admin UI. Custom rules engine and AI solver embedded inside the existing Java/Spring monolith — no rewrite, audit trail intact.

Outcome. Three years live on the trading floor. Drafting time halved. Spring AI bridge before the term existed.

Stack: Java 21 · Spring Boot · custom rules engine · AI solver · Oracle

Recent self-built work

Solo, notebook to production.

Two recent builds demonstrating current capability on the modern stack — from blank repo to deployed Kubernetes, with full observability and measured costs. Built to be run, not demoed.

$0.02per session

60%cache hit

10 weeksto production

Problem. Most LLM projects ship as demos and stay there: no repeatable evals, no observability, no cost ceilings, no fallback when the model is wrong.

What I built. A live cybersecurity strategy game where you play CTO defending an AI-generated fintech against an adversarial agent. Three LLM agents through LangGraph, RAG over five FAISS knowledge bases, semantic caching, deterministic game engine for compliance-sensitive logic. Auth at the edge via Cognito SSO. Full Prom/Grafana/Loki observability.

Outcome. Production-grade agentic AI at near-zero marginal cost — token/inference cost measured per session, not estimated. $160/month always-on, $0.50/month idle, auto-scales to zero.

Stack: LangGraph · FAISS · FastAPI · React · AWS EKS · Terraform · ArgoCD · Cognito

Live → Source → Full breakdown →

79×latency gain

97.5%accuracy

~$280/mooperational

Problem. Most fraud systems catch problems hours too late. Built as a sandbox to stress-test streaming architectures across three industries simultaneously.

What I built. Gaming exploits, fintech transfers, ecommerce checkout — all through one Flink topology. Sub-second alerting with ML inference inline. Airflow DAG handles automated retraining and model promotion.

Outcome. 10K+ transactions per minute, 97.5% accuracy, $300/month operational cost. The interesting part: a 79× performance fix from rethinking the inference pipeline (single-record async → batched sync), not throwing more servers at it.

Stack: Kafka · Flink · Spark ML · Scala · Python · Airflow · Terraform

Source → Full breakdown →

Earlier work

Foundations.

<3 secqueries

Yearsof history

On-demandreal-time

Problem. Product teams needed to slice player data across any dimension — by game, region, cohort, over years of history. Existing tools couldn’t keep up. Every ad-hoc question turned into “we’ll get back to you tomorrow.”

What I built. A custom query engine translating plain requests into optimized SQL across billions of events. gRPC-flexible interface for downstream services.

Outcome. Sub-3-second response on years of data. Product teams unblocked.

Stack: Scala · Akka · Cats Effect · gRPC · custom SQL engine · AWS

×4conversion

Full GTMfrom scratch

EU-compliantpayments

Problem. Connecting dancers with teachers across Europe — fragmented market, no aggregator, manual booking everywhere.

What I built. Marketplace platform plus the operational layer behind it: web scrapers feeding a live event map, automated Facebook posts, A/B-tested funnels. Full legal stack including EU-compliant payments and lawyer-reviewed contracts.

Outcome. 4× conversion rate vs initial baseline. Full go-to-market shipped solo as founder.

Stack: Python · Django · PostgreSQL · Stripe · web scraping · analytics funnels

Let’s talk

A 20-minute call to find your entry point.

No prep, no slides. We locate where you are on the path — diagnose, prove, ship, or direct — and if there’s no fit, you walk away with one concrete suggestion. That’s the offer.

Book a 20-min call →

See the four engagements →

AI, Metrics & Production Data Streaming, Analytics & Pipelines

Real systems. Real constraints. Real numbers.

Production systems, years live.

ESL — Tournament Automation

Société Générale CIB — Termsheet Generator

Solo, notebook to production.

RansomRampage

PatternAlarm

Foundations.

ESL — Analytics Engine

HopGang — Dance Marketplace

A 20-minute call to find your entry point.

You Missed

9 AI Approaches Benchmark for Streaming Classification on Kafka

Streaming Data + AI: Mix ML and GenAI Without Burning Tokens

Secrets Behind a $0.02 Multi-Agent AI Game — From Notebook to EKS

Social Engineering Detection: The Brutal Truth About Static Security and The AI Cure