PatternAlarm
Real-time fraud detection at scale
10K events/min · 97.5% accuracy · 79× faster on the same hardware · ~$280/month
Solution Architect
Real-time data and AI for the systems that can’t afford to fail.
I design and troubleshoot mission-critical streaming architectures, often AI-augmented. Mines generalist engineer, 15 years across fintech, telecom, marketplaces and gaming — I target the 20% of effort that delivers 80% of immediate impact. No endless refactors.
100K+ events/sec · 97% accuracy · 18× speedups · €10M+/day in production
Mission-critical streaming
Real-Time Data, Done Right
Streaming pipelines for fraud, payments, trading, and live ops.
Real-time systems break in expensive ways: lost transactions, missed fraud, SLAs breached at 3am. I design streaming pipelines that hold under pressure — and modernize the legacy Java that surrounds them, without rewrites. Sub-second latency, multi-site resilience, predictable costs.
What I bring to the table
- Streaming pipelines sized for the workload — real-time event processing with an optional analytics layer on top: fraud detection, product telemetry, leaderboards, fault detection. p50/p99 measured under load, capacity that holds during spikes.
- Cost engineering with trade-offs — same throughput on less infra; daily/monthly cost estimates based on your traffic.
- Legacy modernization, two modes — Strangler Pattern for incremental cuts, or full rewrite covered by non-regression tests. No big-bang risk either way.
- Observability that wakes you up for the right reasons — alerts that mean something, runbooks your team can actually follow.
Société Générale CIB
€10M+/day in equity derivatives, 3 years live trading floor
Java/Spring monolith · custom rules engine + AI solver · zero rewrites
ESL Gaming
5 years streaming for global esports tournaments
Multi-region · sub-second · 100K+ events/sec under live ops pressure
For Fintech CTOs · Heads of Data · Banking compliance leads · Live-ops teams
AI Integration
AI That Survives Production
From prototype to deployed system, with measured costs and audit trails.
Most LLM projects ship as demos and stay there: no observability, no cost ceilings, no fallback when the model is wrong. I deploy multi-agent systems with the boring 80% production needs — auth, audit, cost discipline, runbooks — solved before launch.
What I bring
- The right tool for the job — RAG with vector DBs, semantic caching, classical ML with Spark, graph algorithms and rule-based systems, or LLM agents — picked on fit, not hype.
- LLM costs measured per session, per query — unit economics you can show the CFO, not "should be cheap".
- Production AI from 10 weeks — notebook to deployed Kubernetes, solo. Spring Boot or FastAPI microservices, chosen case by case to fit your existing stack.
- Audit-ready by design — rules engines for compliance-sensitive logic, AI for context, structured outputs.
RansomRampage
Multi-agent simulator, notebook → AWS EKS in 10 weeks solo
$0.02/session · 60% cache hit · auto-scales to zero · ransomrampage.com
ESL Tournament Automation
RAG production at gaming scale, 6 months
4-microservice architecture · $0.30/process measured
Société Générale CIB
AI solver inside Java/Spring monolith, 3 years live
€10M+/day · audit-ready · Spring AI bridge before the term existed
For AI/ML Heads with cost burn · CISOs with audit-trail needs · Heads of Data shipping AI products · Compliance officers
Engagement style
Start with the problem.
Engagements usually begin with a specific problem: a pipeline burning cash, a prototype that needs hardening, a Java system that needs real-time data, a platform where downtime cascades into revenue loss. If that’s where you are, the path is short — send a message at aurelien@scalefine.com or book 20 minutes via Cal.
What you get from a first conversation: a read on what’s actually broken, what’s worth fixing first, and whether I’m the right person for it. If I’m not, I’ll tell you and point you somewhere useful — that’s part of the offer.
Delivery style
Prototype → production. Methodology, not magic.
Two real engagements, same playbook: smallest skateboard first, prod-shaped early, measure before scaling.
RansomRampage — 10 weeks, prototype → production
CTO crisis simulator. Three LLM agents around the table — CISO advises, SRE optimizes, the hacker attacks. Up to 20 turns to survive a live ransomware siege on your AI-generated fintech.
| Week | Move |
|---|---|
| 1 | Game design + tech arbitration. Defined rules, turn structure, win/lose conditions, UI layout. Compared different LangGraph + RAG strategies. Chose vector DB, observability stack, microservice boundaries. |
| 2 | 5 domain knowledge bases. 70 FAISS chunks: MITRE ATT&CK, SRE patterns, offensive techniques, fintech archetypes, tech corpus. BGE-small over BGE-M3 (17x smaller). |
| 3–4 | 3-agent LangGraph pipeline. Gateway → cache → RAG → generate → update. Semantic caching at cosine > 0.9999. Structured Pydantic output. |
| 5 | Deterministic game engine + API. Revenue, compliance, breaches resolved by Python, not AI. Hacker queued this turn, resolved next. FastAPI backend. |
| 6 | React frontend + first playtest. 1 user, 9 issues, 3 critical. 23 fixes in 1 commit. Added 20 adversary personas including Kevin from IT. |
| 7–9 | EKS production deploy. 6 Terraform modules. ArgoCD GitOps. CI ~11 min. Cognito SSO at ALB level — zero auth code. |
| 10 | Observability + launch. Prometheus, Grafana, Loki (256MB vs 2GB ELK). $0.02/game, $160/mo, $0.50 idle. |
PatternAlarm — Real-time fraud, 8 weeks solo
Multi-domain fraud detection on live streams. Gaming exploits, fintech transfers, ecommerce checkout — 10K events/min through one Flink topology, sub-second alerting with ML inline.
| Step | Move |
|---|---|
| 1 | Concept + target outcomes. Multi-domain fraud detection. Goals: sub-second alerting, 10K+ events/min, under $300/mo. |
| 2 | Fraud pattern research. Studied velocity abuse, account takeover, payment manipulation. Built synthetic data generator. |
| 3 | Model prototyping in notebook. Scikit-learn gradient-boosted Random Forest. Feature engineering across 3 domains. 97.5% accuracy. |
| 4 | Local env matching prod. Real Kafka, PostgreSQL, Flink via Docker Compose. Evaluated MSK vs self-managed. |
| 5 | End-to-end data flow. Generator → Kafka → Flink + model serving → alerts dashboard. Schema validated. |
| 6 | AWS deploy + training pipeline. Terraform modules. Airflow DAG for automated retrain + promote. |
| 7 | Bottleneck diagnosis. Latency spike under load. Root cause: single-record async inference, not capacity. |
| 8 | Batched sync inference. Same hardware: 79× latency, 59× throughput. Scales to 35K predictions/min. |
| 9 | Open-sourced. Terraform + logic + performance analysis on GitHub. Under $300/mo. |
skateboard first → prod-shaped local → cloud per service → measure before scaling → ship lean
Philosophy
Skateboard before scooter, scooter before car.
Most projects fail because they over-build before knowing what users actually need. Ship the smallest version that works, put it in front of one real user, fix what breaks. Scale comes later — only after the thing is proven worth scaling.
Three principles I apply on every engagement:
1. Build the skateboard.
Whatever proves the riskiest assumption, in the smallest form, with one real user touching it within weeks — not months.
2. Production-grade ≠ over-provisioned.
Every resource earns its cost. Caching added when traffic justifies it, not before. Smaller models when retrieval quality matches. Single instances when one handles the load. Over-sizing is paid every month; right-sizing is measured first.
3. Deterministic where it counts.
AI handles context and language. Plain code handles money, compliance, and anything that needs to be reproducible. Rules engines for decisions that get audited; LLMs for the parts where ambiguity is acceptable.
15 years of finding the 20% of work that moves the needle.
Stack
The toolbox below is here for engineering leaders who want to verify fit before a call.
Domains served
Fintech & Capital Markets · Banking · Telecom · Gaming & live ops · Marketplace platforms · Public sector / GovTech
What people say
Exceptional problem-solving abilities. His capacity to quickly understand complex technical systems and develop creative, unconventional solutions was remarkable.
Toujours à l'écoute, il sait s'adapter aux différents changements et trouve toujours une solution pour obtenir le résultat escompté.
Let’s talk
Mission-critical or production AI? 20 minutes.
No prep, no slides. We figure out if there’s a fit — and if not, you walk away with one concrete suggestion. That’s the offer.