Leveraging AI to Supercharge Development
Production agentic systems, multi-region consumer infrastructure, and spec-driven AI-native development — built from first principles and shipped as live side projects.
Tech stack organized by who builds it. Each lab page enumerates the current tech surface — models, platforms, agents, dev tooling — with cross-references into the architectural deep dives below.
Azure AI Foundry, Azure OpenAI, GitHub Copilot, Responsible AI Toolbox.
Vertex AI (Model Garden, Endpoints, Pipelines, Vector Search, Agent Builder) + Gemini multimodal.
GPT-4o, GPT-5, o-series reasoning, Responses API, Realtime, Codex CLI, Swarm agents.
Claude (Opus / Sonnet / Haiku), Claude Code, MCP, Computer Use, Constitutional AI.
DGI is here — AGI’s practical cousin for engineering. A multi-model LLM council embedded in the architecture view: interrogate any component live, and the spec crystallizes underneath. Architecture as the Product.
A real agent built end to end — requirement to evals — with every decision on the record. The model decides; a deterministic harness executes. Nine single-concern files, the injection defenses, the money gate, and a golden set with a negative control. Screenshots of the actual IDE and process.
First-hand friction log evaluating open-weight models on an M4 Max / 128 GB laptop — GLM-4.5-Air in MLX-4bit, the memory math that rules out the 750B models, Ollama vs LM Studio, and why the honest verdict is a frontier API with a provider-agnostic hedge.
Three-layer agentic framework: single-agent loop, multi-agent orchestrator, modular tool SDK.
Multi-region typing platform: voice narration, 8 languages × 10 keyboard regions, anycast routing.
AI content factory: spec-driven, RAG, Report Cards, model routing, self-correction.
Independent YouTube archives — every full session on its own subdomain, six languages, refreshed nightly. One template; a new archive is a persona file.
Coding agents shortened the build; they did not shorten the impact analysis. One business ask followed to production twice — the traditional path, counted, and the path where one owner drafts the brief, impact map, API contract, mock and stories before the first meeting.
A property-insurance POC — built on years of group-benefits platform experience — showcasing agentic systems around the full pipeline: auto quote generation, configuration-driven products, and an underwriting console where AI drafts and a human owns the decision.
The complete free study guide for Anthropic's Claude Certified Architect exams — 5 domains, real-payload walkthroughs, 50+ glossary cards, deployment architectures, capstone.
A diagram-led, first-principles explainer of how frontier labs run personal AI at scale — the context pack, memory namespaces, the stateless fleet and its control plane, token economics, quality and trust — and the one-to-one mapping onto enterprise systems.
The managed AI service layer for the post-hyperscaler era — vendor-agnostic, infra-first, BYO-everything. 5 pillars + control plane + 4 deployment envelopes.
Personal RAG, 100% local by default. Drop a folder, ask cited questions from laptop or phone (Telegram bot). MIT-licensed, v0.11.
Founder runway dashboard: bank statements in → sliding-window burn rate, runway, and a decision deadline out. Built Mar 2026, used daily since.
How I think about LLMs, RAG, agents, and production AI — first principles to shipped systems.
Full production RAG for real verticals — HR, customer transactions, healthcare. Reference architectures end-to-end.
LangChain, CrewAI, AutoGen, LlamaIndex, Pydantic AI — honest comparison with code, pros/cons, and when to pick each.
The open-source low-code canvas for LLM apps and agents (acquired by Workday for their enterprise AI agent platform). Hands-on series — install, build, and database-level observations as I work through real flows.
Wrap your existing REST API as an MCP server in ~60 lines of TypeScript. Three surfaces — REST for developers, MCP for AI agents, CLI for power users — sharing one source of truth. Runnable subscription-status example.
An email inbox rebuilt as a production system — Gmail API, SQLite sender catalog, 4×/day cron, human-in-the-loop over Telegram, Claude Code as the interface. Sender-first triage instead of email-by-email.
The four routing patterns I use in production: rule-based, classifier-based, cascading, parallel adversarial. With the LLM Council pattern — Claude orchestrates, Gemini reviews, Codex validates — for high-stakes tasks.
Your AI vendor is a dependency, not a destiny. Six layers decomposed — the provider gateway, the model-selection decision matrix, guardrails, multi-tenant isolation, and encryption in transit (mTLS) and at rest. Runnable code, animated data-flow diagram.
Langfuse, LangSmith, Braintrust, RAGAS, DeepEval, LlamaGuard — how the industry is adopting LLM observability, evaluation, and safety tooling. Real code for each.
Control plane + data plane + guardrails + observability. Vendor-neutral reference with Microsoft / AWS / custom stack mappings.
A career across Java, Python, TypeScript, SQL, NoSQL, Kafka, Spring Boot, React, Next.js — and why stack breadth is a multiplier in the AI era.
The production deep dives above represent my core AI-native work. The tabs below cover adjacent areas I actively explore: Onyx RAG prototyping, MCP server patterns, Claude Code excellence techniques, and polyglot AI strategy for multi-model orchestration.
Retrieval-Augmented Generation Systems
Model Context Protocol Implementation
Advanced AI-Powered Development
Multi-Model AI Orchestration
As AI capabilities evolve at breakneck speed, staying ahead means mastering not just individual models, but the art of orchestrating multiple AI systems to create applications that are intelligent, scalable, and cost-effective.