Blog
Short, pragmatic notes on production AI — patterns, failures, and fixes.
Start here
Shamir Secret Sharing in Go: Eliminating Single Points of Failure
GF(2^8) Shamir implementation from AxisCorePay in production Go: MPC-TSS vs Shamir 2-of-2, three pitfalls, and crypto settlement architecture.
Why Backtests Mislead: CAGR, Expectancy, Omega Ratio, and SQN
Win rate misleads. Sharpe rewards bad patterns. Arithmetic return overstates real gains. The 4 metrics production algo trading uses instead — with live calculators.
Case Notes: MTRobot — Building a Multi-Tenant Algo Trading SaaS on MT5
Engineering notes on building a multi-tenant algo trading SaaS — MT5 integration, execution isolation, real-time streaming, and security constraints. What we learned building production trading infrastructure.
Case Notes: RoomIQ — Teaching AI to Place Furniture
How we built RoomIQ's Interior Engine — a hybrid system combining LLMs, deterministic geometry algorithms, and evolutionary optimization to generate physics-valid 3D room layouts from natural language.
Model Skewing in Production: What It Is, Why It Happens, and How to Fix It
PSI thresholds, KL divergence, and a 7-step debugging workflow for detecting model skewing, data drift, and training-serving skew in production ML systems.
Case Notes: Steve Trading Bot
From backtests to controlled live rollout in trading systems.
Case Notes: AgrigateVision
Lessons from deploying agriculture CV under variable conditions.
RAG Architecture Patterns for Production: Chunking, Hybrid Search, and Access Control
How to build RAG that works in production: chunking strategies, hybrid search, RBAC/ABAC access control, evaluation, and latency optimisation. Visual guide.
Why Machine Learning Models Degrade in Production: 5 Failure Modes
Why ML models degrade after deployment: data quality breakdowns, pipeline drift, monitoring gaps, ownership failures, and training-serving skew - plus a practical debugging workflow.
Applied AI Is Not a Web Service: Why AI Projects Fail After Deployment
Why AI projects fail after deployment: treating AI like a web service. What changes in architecture, monitoring, ownership, and delivery.