Jared Frieden · AI Systems Architect & Engineer
AI Systems Architect & Engineer

I design agent systems, and I use them to ship real software.

I'm Jared. I built an operating system out of AI agents and put it to work shipping real software. The products, tools, and games below all came out of it. I designed the system, and I built what it ships.

16 working prototypes 440+ tasks orchestrated 83 batches run 145 curated memories ~80 min idea to prototype
Featured work

Three I'd put in front of anyone.

Rust daemon hub · routerworkflow engine Local models first in the chain Cloud providers bounded failover MCP tools client spokes Desktop shell Theia-based SQLite audit chain versioned events · replayable route fallback
Rust · local-first
01

PocketBridge.Ai

A local-first AI gateway in Rust that grew into a full agent platform: provider routing, SQLite persistence with an audit chain, and a Theia desktop shell. Tagged v0.1.0 and release-ready, held behind a dogfooding gate until it clears one real task end to end.

v0.1.0 · gatedRead the case study →
Studio browser editorAR + QR campaigns Phone live AR viewno app install Scan telemetry every scan, locatedin space Voxel heatmap engagement, rendered in 3D ~2s push WebSocket scans aggregates informs the next edit
React · three.js
02

Mango-QR

A browser platform for building AR and QR campaigns and measuring real engagement. Studio edits reach a phone in about two seconds over WebSockets, and scan telemetry renders as spatial voxel heatmaps.

working demoRead the case study →
Photo what did the pet find? Vision model identifies the itemnever judges it Toxicity database curated · citedthe only verdict source Invariant test no source →never returns "safe" Verdict with its citation item name verdict passes
Cloudflare · Expo
03

SniffSafe

A pet-toxicity scanner deployed on Cloudflare Workers. The vision model only identifies the item; every verdict comes from a cited database, enforced by an invariant test that blocks any path from returning 'safe' without a source.

live API · prototypeRead the case study →
The engine

How one person ships this much.

Four systems do the heavy lifting. They plan the work, route it, build it, review it, and keep track of what happened, and I stay the one making the decisions. This is the reason the catalog below exists.

Human decisions, gates,course corrections Orchestrator — Claude Code framing · planning · dispatchquality gates · memory Executor CLIs implementation lanes Media / research lanes bounded fan-out Vault — message bus task files + frontmatterfolder = lifecycle state Daemon deterministic state machineno LLM in the loop Review swarm tiered: auto / agent / humanverdicts gate closure Persistent memory file-based, git-backedrules · projects · workarounds brief / sign-off writes TASK-*.md claim & execute deliverables recall / persist moves files on state change verdicts solid = agents and state · dashed = knowledge · amber = deterministic machinery and gates

Agent Operating System

A hub-and-spoke orchestration layer: one architect model routes work to executor agents through a plain-markdown message bus, with deterministic scripts holding the state. It stays human-readable and version-controlled, and survives any single tool failing.

The Factory

An idea-to-prototype build engine. A nine-question interview locks a spec, then the pipeline designs, builds, verifies, and delivers autonomously. A proof run went from brief to a running web and mobile app in about 80 minutes.

Model Routing Economics

A measured cost-control policy: frontier models plan and judge, mid-tier models execute, and scripts handle everything deterministic. Daily telemetry tracks where the spend actually goes.

Persistent Memory & Curation

A file-based memory that survives context resets, with freshness contracts, automated drift sweeps, and a git-backed curation dashboard. It is audited and corrected like production data.

The rest of the work

Everything else, with an honest status on each.

how done is it?
Products

Mango-QR

A browser platform for building AR and QR campaigns and tracking engagement, with a two-second studio-to-phone live push and spatial telemetry heatmaps.

React · three.js · FastAPI working demo

Jira Insights

Natural-language questions over Jira delivery metrics, running as a single Cloudflare Worker. SQL is validated by an AST parser, after the previous string-matching guard proved bypassable in testing.

TypeScript · Cloudflare Workers · D1 built · pre-deploy

Headlands & Co

The marketing site and Python lead-scoring pipeline for my consultancy, deployed on Cloudflare Pages.

Node · Cloudflare Pages · Python live

tip-split

A no-account restaurant bill splitter, and the app The Factory built end to end to prove the pipeline.

Cloudflare Workers delivered
Agentic systems & tooling

Ostia-Aevum

A remote MCP server that serves agents pre-validated building blocks. Live on Cloudflare Workers.

TypeScript · Cloudflare Workers · MCP live

memdash

A zero-dependency dashboard for agent memory: a force-directed graph of the knowledge base and a curation queue for stale entries. Built on Node's native SQLite.

Node · node:sqlite in daily use

aseprite-mcp

An MCP server that lets an agent create and edit sprites by driving Aseprite's CLI through tool calls.

TypeScript · Node · MCP working tool

Council

A multi-persona deliberation engine that runs a real cross-examination and returns the points of agreement and the open disagreements.

Claude Code · Obsidian in use

Delivery Highway

A hard-gated shipping pipeline with app-agnostic gates and an append-only delivery ledger, so nothing reaches users before it has earned each stage.

PowerShell · Pester phase 1 live

Thinking-Tools

Three custom reasoning skills that auto-invoke on context: divergent ideation, a structured decision ladder, and fast topic ramp-ups.

Claude Code in daily use
Games & interactive

Come to Life

A standalone game with an AI-generated, painted-on-3D art pipeline, ported from Godot to Unreal Engine 5.7 with its design intact.

Unreal Engine 5.7 · C++ greybox

runeforge-frontier

A browser-based 3D city-builder with switchable first and third-person views, built on Three.js.

Three.js · Vite early prototype
Systems programming

UniversalControler

A Rust daemon that maps a game controller to full Windows control: pointer, keyboard, T9 text entry, and agent dispatch.

Rust · Direct2D WIP
Explorations

iOCT

A blueprint for a Telegram-native operations and analytics tool, shown as a design exploration rather than a shipped product.

Python · Telegram concept

Plus the operational tooling behind all of it: a portable, secret-free backup of the agent "soul," and shared git-hook gates enforced across every repo.

Doctrine

Rules the system runs on.

These are the rules that let one person ship a lot without quality quietly sliding. Each one started as a real mistake, got written down, and is now enforced by a script or a review step. Here is how they fit together, then what each one does.

Frame it plan beforeyou build Build it make the thing Verify it checks, evidence,then a person Prove it one real task,end to end Ship it claim only whatyou can show the ratchet: every lesson raises the bar, nothing lowers it

Frame before you build

Every big or hard-to-undo move gets one pause first: name the real goal, sketch the approach two different ways, and keep that plan visible so the work does not drift off course as it goes.

Verify in stages

Whether something works gets settled in stages: automated checks, then hard evidence, then a human review. When quality comes down to how something looks or feels, a person makes the final call, never an AI on its own.

Floor and ceiling

Quality sits between two marks: a checklist sets the minimum to pass, and a concrete example sets the target to reach. Criteria alone drift toward "good enough," so an example of "great" is always pinned beside them.

Prove it, don't just build it

Finishing the code is not the same as it working. Before anything counts as done or reaches a user, it has to clear one real task from start to finish. Passing tests alone is not enough.

The ratchet

Standards only move one way: up. Every lesson learned raises the bar and gets recorded, so a later shortcut cannot quietly lower it. The system gets stricter over time, never looser.

No hype

Quality is never claimed until it has been checked directly. State what actually happened and point to the proof, instead of reaching for adjectives. Whatever has not been seen working does not get called good.

About

I'm Jared Frieden, an AI systems architect and engineer. I run a one-person software practice that works more like a small team: the systems on this page plan the work, build it, review it, and remember it, and I stay the one making the decisions. I'm looking for a role at that level, designing and building agentic systems and the tools around them.