Build Notes

What it does. A governed AI knowledge workspace that turns scattered documents, databases, and institutional memory into cited answers and safe automated actions. Teams get answers they can verify, plus agentic workflows whose actions pass through approval gates before anything happens.

Complexity. Very high. It combines multi-tenant workspaces with RBAC, two-stage retrieval with calibrated cross-encoder re-ranking, governed organizational knowledge with human review queues, and agentic workflows with approval gates, and each of those is a hard problem on its own before they have to coexist.

Experience it took. Retrieval engineering and re-ranking, RBAC and multi-tenant security modeling, governed-knowledge workflow design, agent tooling over MCP with OAuth-gated actions, and model-agnostic LLM integration spanning local and hosted providers.

Read more about MightyRAG

What it does. A CRM and manufacturing workflow system covering the full path from lead and quote through measurements, production, and installation to invoicing and payment. A formula-driven pricing engine quotes doors and railings in seconds, and it is live in production at prodoormalta.com.

Complexity. High. The formula-driven pricing engine with its price matrix is genuinely intricate, three applications (backend, internal ERP, public site) share one typed contracts package, and the whole thing runs as a real Azure production deployment with a runbook, which is a different discipline from making a demo work.

Experience it took. PostgreSQL data modeling, pricing-engine and formula-matrix design, TypeScript contract-first development with Zod, OpenAPI, invoicing and payment flows, and Azure production operations on App Service and PostgreSQL Flexible Server.

Read more about ProDoor

What it does. Project and task management where every task is executed by an AI agent, with the work streamed to a live, append-only audit trail. Per-project knowledge lives in PostgreSQL, and delegation becomes something you can inspect line by line.

Complexity. Very high. Running an agent per task inside a multi-tenant system means the credential, sandboxing, and egress controls have to be right, and the append-only audit trail streamed live over SSE has to stay trustworthy under exactly the conditions where an agent misbehaves.

Experience it took. Multi-tenant security modeling (credentials, sandboxing, egress), agent runtime work on the Strands Agents SDK, PostgreSQL and pgvector, SSE streaming, and a Deno and React application stack.

Read more about SuperConductor

What it does. A wellness-business platform handling online bookings synced with Google Calendar, a services catalog, loyalty rewards, payments, owner analytics, and AI-assisted follow-up plans, with automated email flows so practitioners can stay present with clients.

Complexity. Moderate. The individual patterns are well-trodden full-stack territory, but there are a lot of moving parts to keep honest: two Google API integrations, a Redis-backed email queue, live updates over SSE, and AI features designed to degrade gracefully when no model is available.

Experience it took. React, Express and Prisma, Google Calendar and Gmail API integration, Bull and Redis queue workers, SSE live updates, payments and loyalty logic, and LLM integration via OpenRouter with graceful degradation.

Read more about RAY Healing Center

What it does. Our software-delivery method, productized: a self-contained AI delivery team of 35 coordinated skills that design, build, review, and ship software through nine pipelines with evidence gates at every stage. It is the discipline behind every other system on this page.

Complexity. High. There is no server to deploy, but designing 35 skills that coordinate through machine-readable gate and pipeline contracts, keeping validators deterministic, and making the whole thing run on three different agent hosts is systems design of a rarer kind: the system being engineered is the process itself.

Experience it took. Delivery-process design, skill and prompt engineering, deterministic validation tooling, machine-readable contract design in YAML, and portability work across multiple agent platforms.

Read more about Fabled

What it does. One self-hosted gateway for all your LLM traffic: multi-provider routing, virtual keys, team budgets and rate limits, exact and semantic response caching, and MCP tools, all behind a single OpenAI-compatible endpoint with a React control plane.

Complexity. Very high. A 23-item capability map in a single process is a lot of surface: correct budget accounting and rate limiting across providers, semantic caching over pgvector, and acting as MCP client, MCP server, and inference gateway at once, all validated by a measured benchmark methodology rather than by optimism.

Experience it took. Deno, multi-provider LLM API integration, PostgreSQL and pgvector semantic caching, rate limiting and budget accounting, MCP protocol work on both the client and server side, React, and benchmark design and measurement.

Read more about Frosty Gateway

What it does. An MCP server with a real, sandboxed Chromium browser inside it. AI agents drive the web through 19 deterministic tools while a human can watch a live screencast and take over at any moment, and persistent named sessions keep logins alive without credentials ever passing through the model.

Complexity. High. Sandboxing a full browser in a slim container, streaming a live view with mouse, keyboard, and IME input under an exclusive takeover lock, and layering hashed API keys with an optional OAuth 2.1 resource-server mode is security and systems work, not glue code.

Experience it took. MCP protocol work, Chromium automation and sandboxing, container image hardening, live streaming and input handling, API-key and OAuth 2.1 security design, and Deno.

Read more about Browser-MCP

What it does. A local-first memory server that gives AI agents a persistent, searchable long-term brain over the Model Context Protocol. Memories are stored as vectors in your own PostgreSQL, isolated by namespace, with a bring-your-own-model design that runs fully local by default.

Complexity. Moderate. It is the earliest-stage system of the ten and deliberately simple: a FastAPI application over PostgreSQL and pgvector with namespace isolation instead of a heavyweight auth stack. The judgment call about what to leave out was most of the design work.

Experience it took. Python and FastAPI, PostgreSQL and pgvector, MCP integration, and local-first LLM endpoint design against OpenAI-compatible APIs.

Read more about Open RLM Memory

What it does. A Docker-Desktop-style Windows app for WSL containers, shipped as a single compiled executable with no runtime dependencies. Every button maps to a real, documented CLI command, and it imports docker-compose and Kubernetes definitions into ordered run plans.

Complexity. Moderate. The honest-tooling stance keeps it disciplined rather than sprawling: capability gating means features exist only when the host provably supports them, and the compose-subset compiler that turns stack definitions into ordered CLI plans is the one properly tricky piece, all held together by a 166-test suite.

Experience it took. Windows desktop development with Deno, React, and WebView2, CLI integration and capability detection, compose and Kubernetes import translation, and disciplined test engineering.

Read more about WSL Container Manager

What it does. A sovereign, high-assurance data platform that replaced a fragile spreadsheet estate with governed, auditable market-intelligence data. A promotion gate controls what becomes trusted data, and native Office document generation delivers workbooks, reports, and presentations from pinned data snapshots.

Complexity. High. The security model runs deep (Argon2id with pepper, RBAC, database least-privilege, sandboxed workers), the governed promotion gate makes data trustworthiness an engineered property rather than a hope, and the formula engine's "NULL is never 0" rule is the kind of correctness obsession that only matters until the day it really matters.

Experience it took. Security modeling (Argon2id, RBAC, least-privilege), PostgreSQL, sandboxed worker design, governed data-promotion workflow design, native Office Open XML document generation, formula-engine correctness, and test-coverage discipline at a 95 percent gate.

Read more about Astutus Data Manager