Designed a consistent adapter/factory architecture (getAI(), getDatabase(), getCache(), …) applied uniformly across all 31 packages, isolating business logic behind stable interfaces — swapping a managed API for a self-hosted alternative is a config change, not a refactor, eliminating vendor lock-in and de-risking provider price/API churn.

Unified 40+ third-party providers behind those interfaces, including LLMs (OpenAI, Anthropic, Gemini, AWS Bedrock, Hugging Face), databases (PostgreSQL, SQLite/Turso, DuckDB, with vector search), caches (Redis, S3, memory, file), payments (Stripe, BTCPay, USDC/x402), and auth (Keycloak, Cognito, Nostr).

Built as a higher-level alternative to early LangChain.js (then well behind its Python counterpart): one interface for chat, streaming, embeddings, vision, TTS, and function calling that keeps pace with the newest models without framework churn.

Leveraged AI-assisted development to write thin first-party adapters instead of adopting heavyweight frameworks — vendor SDKs are optional peer dependencies, so consumers install only what they use, shrinking the dependency tree, upgrade pain, and supply-chain exposure.

Made the SDK AI-native: every package ships an AGENT.md machine-readable context file plus an MCP documentation server, so AI coding assistants produce correct integration code on the first attempt.

Ran the project with production-grade engineering practice: pnpm/Turborepo monorepo, ESM-only, ~46k lines of Vitest tests, Biome linting, conventional commits, and automated Changesets releases to public npm — 1,800+ commits.

Designed a single-source-of-truth model layer: a @smrt() decorator on a TypeScript class triggers a build-time AST scan (custom OXC-based scanner → manifest) that generates the class's SQL schema, REST API, CLI, and MCP tool server from one definition — so a domain concept is defined once instead of re-implemented as boilerplate across four surfaces, cutting the code an AI agent writes and a human reviews.

Unified the human and agent surfaces onto one code path: generated MCP tools, REST routes, and CLI commands all resolve to the same ORM collection under a shared sensitive/readonly field policy — letting AI agents operate an application with the same capabilities and guardrails as its users, from a single codebase.

Embedded AI operations in the object model — is() for natural-language boolean checks and do() for instructions on every persistent object, with the object's own methods passed to the model as function-calling tools — so domain logic is directly promptable rather than bolted on.

Built ~40 lockstep-released packages of hardened, reusable domain logic — multi-tenant RBAC (a four-level permission cascade), double-entry billing (single-table-inheritance contracts posting balanced journals to a ledger), plus content, messaging, assets, semantic knowledge, and geospatial search — so bespoke SMB and greenfield SaaS builds compose vetted modules instead of re-solving the same concerns per project.

Shipped a typed, themeable, accessibility-tested Svelte 5 component library (~80 components — forms, tables, modals, navigation, calendar) so generated models get a usable UI without rewriting the same components for every app.

Made the framework agent-native and reproducible: per-package AGENTS.md context files and an MCP developer server (code generation + project introspection) so AI coding assistants integrate correctly on the first try; run as a pnpm/Turborepo ESM monorepo with hundreds of Vitest test files, Biome, conventional commits, and automated Changesets releases to npm.

Launched a network of 13 hyperlocal news sites (e.g., bentleyalberta.com) powered by a multi-tenant SaaS dashboard that orchestrates a suite of specialized AI agents sharing a unified database via Single Table Inheritance (STI).

Autonomous agent monitoring municipal portals (CivicWeb/SharePoint) using PDF parsing and OCR to generate source-linked meeting coverage and automated journalism.

Managed automated media workflows orchestrating performers, characters, and scenes through complex ComfyUI workflows and custom diffusion models.

Developed agents for automated hyperlocal weather forecasting and sports statistical analysis for real-time content delivery.

Engineered agents for autonomous social media performance tracking/posting and automated GitHub repository management.

Developed a decentralized ad network with weighted distribution, impression tracking, and automated registration workflows.

Architected a TypeScript monorepo (6 apps, 2 agent daemons, 6 packages) around a model-driven framework: 59 declarative domain models generate REST APIs, Postgres/SQLite schema migrations, CLIs, and MCP servers from a single source of truth.

Built a local-first capture system with native device adapters — Sony CrSDK (C++ bindings), gphoto2, Insta360 MediaSDK (360°), and hardware turntable control — where sessions record offline to SQLite + disk, then resume-sync to the central library as immutable, checksummed source packages.

Automated field-media offload (drones, cameras) with content-based file classification, sidecar detection, SHA-256 deduplication, and GPS track indexing with geohash spatial queries that link footage moments to real-world places — the substrate for reconstructing whole neighbourhoods from a single street drive.

Designed a lease-based distributed job orchestrator that schedules each pipeline step by placement — edge capture node, central host, or NVIDIA/ROCm GPU worker — so heavy AI workloads run on hardware I own instead of metered frontier APIs.

Shipped 13 production pipeline step types: ComfyUI image editing and text-to-video generation, Whisper transcription, voice-cloning TTS, face embedding and matching (InsightFace), background removal, and an article-to-video news-segment generator that publishes straight to YouTube/TikTok.

Eliminated cloud warm-storage costs with a multi-store architecture: per-tenant S3/MinIO stores with encrypted credentials, presigned multipart uploads, and node-local NAS replication — the library scales by adding servers, not cloud spend.

Built the multi-tenant consumer surface — capability-scoped API keys, HMAC-signed webhooks, tenant provisioning — and published a typed SDK and embeddable UI components to npm, letting downstream apps use ergot as their asset backend without carrying its weight.

Operated the platform with GitOps: Kubernetes + Flux across dev/staging/production, multi-stage Docker images, environment-gated side effects, and a zero-downtime schema-migration workflow.