Jan 2012 — Present
Scarlet Beast
New York City · Remote · scarletbeast.com
Formerly Scarlet Consulting Corporation (merged)
Founder, Lead ML / AI Engineer & Principal Engineer — Hiss (Poker AI), GROWL (Algorithmic Trading) and client engineering
Scarlet Beast was founded in 2012 as Scarlet Consulting Corporation, a client engineering practice, and the two companies have since merged; the ML platform began in 2017.
Founder and principal ML engineer of Scarlet Beast, building Hiss — a production poker AI — and GROWL, a crypto / forex exchange and algorithmic-trading platform. I own the full ML lifecycle: distributed training infrastructure through to real-time, latency-sensitive inference in money-moving environments. The work spans deep reinforcement learning, computer vision, time-series modelling, and the MLOps that ties them together across heterogeneous, cost-optimised compute.
Hiss — the milestone ladder (documented at nn.scarletbeast.com)
- M0 — Correctness gate. Validated the toolchain where the answer is known: tabular CFR / CFR+ / external-sampling MCCFR to ~0 exploitability on Kuhn and Leduc, then the neural path against the same target. Established that an apparent Leduc plateau was undertraining rather than a ceiling — advantage steps 64 → 250 at lr 1e-2 moved exploitability 0.17 → 0.0719.
- M1 — Card abstraction. Suit-isomorphism canonicalisation (~24× precompute reduction), potential-aware equity-distribution features (E[HS], E[HS²]), and k-means under Earth Mover’s Distance into ~1–2k buckets per street over 169 lossless preflop classes. Measured that solve quality peaks at 24 boards: stratified board selection paid +8.4 bb/100 where simply adding boards did not.
- M2 — Six-max game & traverser. External-sampling Monte-Carlo CFR over the bucketed abstraction with legal-action masking asserted at the boundary, gated on reproducing a published two-player subgame result before any six-handed number was trusted.
- M3 — Deep CFR / SD-CFR training. Advantage networks over reservoir buffers, retrained from scratch each iteration, with the snapshot history serving as the average strategy. Resumable bit-identically across reboots and scheduled under systemd; ~845 s/iteration at 16 threads on commodity CPU.
- M4 — Distillation & serving. Collapsed the snapshot mixture into one supervised policy network and exported it into the existing serving contract — TorchScript plus positional normalisation — as 86 features → 384 → 384 → 192 with policy, bet-size and value heads (259,599 parameters), answering in about a millisecond on a CPU shared with the vision pipeline.
- M5 — Evaluation. Opponents calibrated against 142,000 real hands, common-deals paired comparison, and promotion only on a confidence-interval lower bound. Rebuilt an arena that had been scoring a policy which never played, and a convergence metric that had fitted a champion swap as a trend line.
- M6 — Ten-pocket six-max game. Authored a custom OpenSpiel game wrapping the rules engine with its own action space and information state, after measuring that the stock abstraction could not express a half-pot bet (blueprint mass exactly 0.0000) and that the full abstraction’s tensor could not distinguish a 4-chip bet from a 200-chip one.
Measured outcomes. Fold-discipline work moved the live champion from −32 bb/100 to −4 against opponents calibrated on real hands; the CFR blueprint measures +20 to +28 bb/100 heads-up. Serve-time guardrails — a fold-to-three-bet rail worth +13.3 bb/100 — cover behaviours the network would not hold on its own.
Documentation & source. Full technical write-up — concepts, the milestone ladder, the served architecture and a from-scratch rebuild runbook — at nn.scarletbeast.com. Source: github.com/christianscarlet2/hiss.
Distributed training & MLOps
- Training built on heterogeneous, cost-optimised compute: AWS SageMaker as the managed backbone for large training jobs, hyperparameter tuning, experiment orchestration and model hosting; Oracle Cloud (OCI Ampere A1) and Kaggle GPU / TPU kernels for high-volume self-play data generation on spot and free-tier capacity.
- PyTorch, NumPy, Pandas, scikit-learn, CUDA / cuDNN, Docker, and MLflow / Weights & Biases for experiment tracking and reproducible, containerised pipelines.
Computer vision
- OpenCV for image preprocessing, ROI and object detection, and table-state extraction, with Tesseract OCR reading cards, stacks and bet amounts into structured features that feed the Hiss inference pipeline directly.
GROWL — quant / algorithmic trading
- Extended the same ML platform to live markets: time-series forecasting and signal generation over crypto and forex data (LSTM / Transformer and gradient-boosted models), a bot marketplace for algorithmic strategies, backtesting and walk-forward validation, and live order-book and market-data ingestion.
- Owned the surrounding fintech concerns — execution latency, risk controls, and a unified accounting layer (“The House Ledger”) exposed via REST and GraphQL APIs.
Platform & infrastructure
- Visual C++ high-performance inference engine, Python services, a React + Laravel application surface exposing public REST / GraphQL APIs, an SSO auth gateway unifying every product, and multi-tenant Linux / PHP-FPM hosting across the ecosystem.
- Source: github.com/christianscarlet2/hiss
Client engineering practice (since 2012)
An AI-empowered software consulting and digital infrastructure company focused on scalable web platforms, automation systems, performance engineering and enterprise modernisation.
Blending technical leadership with growth strategy and technical sales, I specialise in helping businesses turn operational bottlenecks into scalable systems — using AI-assisted development workflows, modern cloud infrastructure and full-stack engineering. Fifteen years across enterprise and startup environments, working directly with decision makers to architect and deliver high-performance solutions across eCommerce, SaaS, automation and digital operations.
- AI-assisted software development and automation
- Full-stack LAMP, Node.js, React and cloud engineering
- Magento / Adobe Commerce and Shopify Plus ecosystems
- Technical sales, solution consulting and client strategy
- Performance optimisation and Core Web Vitals improvement
- AWS infrastructure, CI/CD pipelines and DevOps workflows
- API integrations, Stripe payment systems and scalable architecture
- Enterprise migrations and modernisation initiatives