Foundation Operations  ·  Stack & Workflow  ·  How We Build

AI-assisted development with accountable review.

AI tools help with research, implementation, testing, and operations. Access is scoped to the task, consequential actions retain human approval gates, and the result is verified against source, tests, deployment state, and live behavior.

$fo dev --status
● Stack: React · TypeScript · Node · Python · PostgreSQL
● Infrastructure: self-hosted services · documented deployment paths
● AI: scoped tools · human approval gates · verification
Core Stack

The primary technical stack

A stable set of tools used across projects where it fits the problem and the operating environment.

Frontend
React + TypeScript
React and TypeScript support typed application interfaces and shared contracts. Component libraries, real-time interfaces, and state management are selected to fit each project.
React 19TypeScriptTailwind v4VitetRPC
Backend
Node.js + Python
Node.js supports APIs and real-time services. Python supports automation and data-processing tasks. Background job queues are used where work should not block a request.
Node.jsExpressPythonSocket.IOJob Queues
Database
PostgreSQL
PostgreSQL supports relational data models, with vector extensions where semantic search fits the project.
PostgreSQL 16pgvectorRedisDrizzle ORM
Infrastructure
Self-Hosted Infrastructure
Docker supports containerized services on self-hosted infrastructure, with versioning and deployment practices defined per project.
DockerNginxUbuntu LTSWireGuardGitHub
AI-Assisted Workflow

Useful tools with bounded access

AI tools can support research, implementation, testing, and selected operational work. Access is scoped to the task, and consequential changes retain human approval gates.

Custom MCP Server
Task-scoped tools and review paths.
A self-hosted Model Context Protocol server can expose the resources needed for a defined task. The operator reviews access, retains approval for consequential actions, and verifies the result.
Source ReviewTestingDeployment StateLive VerificationApproval Gates
Model Router
Right model for the right task.
A self-hosted proxy can route requests by complexity — faster models for routine tasks and more capable models for harder problems. Routing remains a tool choice, not a substitute for review.
Request RoutingLocal InferenceMulti-Model
Session Context
Context that survives sessions.
Project-specific architecture notes and closeout records can carry decisions between sessions. They provide useful context while source and current state remain authoritative.
Master MD FilesSession LogsArchitecture Docs
Search Layer
Selected web context.
A search layer can collect selected web results for review before material enters a model context.
Search APIContext Review
Philosophy

How the work gets done

These are the current operating principles applied according to the risk and needs of each project.

01 ──
Version the work that runs.
Keep running code and operational scripts in source control with documented change paths so work can be reviewed and reproduced.
02 ──
Fail fast. Don't mask errors.
Self-documenting code, clean naming, comments only where logic is genuinely non-obvious. Let it throw. Swallowed errors become production mysteries.
03 ──
Design reusable boundaries.
Shared architecture and configuration boundaries reduce unnecessary rework while leaving room for business-specific workflows.
04 ──
Own the critical path.
Self-hosting is considered when control over data, infrastructure, or cost warrants it. Vendor services remain explicit engineering choices.
05 ──
AI is a force multiplier, not a magic button.
Structured context, scoped access, review, and verification make AI assistance useful without transferring responsibility for the result.
06 ──
Choose tools according to the task.
Search, model selection, deployment, and other supporting tools are chosen according to the project's needs; responsibility remains with the developer.
Mental Models

How AI actually works

These concepts explain how context, routing, and tool coordination shape the workflow.

Context Windows
Model calls depend on supplied context.
A model does not independently retain the conversation between calls. Applications may resend, summarize, cache, or otherwise manage prior context, so the behavior depends on the implementation.

This is why context management matters. As the supplied context grows, less room remains for the current work. The useful material should be kept relevant and within the model's context window.
Context has to be supplied and curated. A growing transcript is useful history, but it can also crowd out the material needed for the task at hand.
You →
hi
Model →
hi · hi
You →
hi · hi · how's your day?
Model →
reads all of it → good.
Model Routing
Route by hardness, not by habit.
A model router can classify requests by complexity and direct them to different configured models.

Routine requests may use a smaller model or local inference, while complex work may use a more capable model. The selected output still requires review appropriate to the task.
How many dimples on a golf ball? → Local · free
Summarize this log file. → Fast · cheap
Write a TypeScript migration for this schema. → Mid-tier
Design the architecture for a multi-tenant platform. → Heavyweight
↳ request classified · configured model selected · output reviewed
Orchestration
Direct the system and stay accountable.
AI-assisted work can combine a model with selected supporting tools instead of asking one system to perform every task.

Search can be handled by a dedicated API, routing by a proxy, and selected tasks by local inference. You define the goal, review the output, and approve consequential steps.
AI tools can multiply implementation capacity, but the operator still owns the decision, the review, and the result.
You
Define the goal. Direct the work.
AI Model
Drafts and implementation assistance.
Fast Model
Routine task assistance.
Search API
Selected web results.
MCP Server
Task-scoped tool access.
Local Inference
Local model option.
Search Context
Review search context before use.
Raw-page retrieval can include navigation, markup, and unrelated text. A search API may return selected excerpts for model review.

The underlying source still needs to be inspected for important claims.
Source Page
Page content and navigation
source material
→
Search API
Returns selected excerpts
candidate context
→
Source Review
Inspect important claims
reviewed context
Treat search output as context to review, not as authority.