ECHO Projects

Member tooling · Digital Orukami Launched

SIGNET

A signature studio in one file. Type your name in a script face, sign on a pad that follows your pen speed, or photograph a real signature and let the tool lift the ink off the paper. Everything lands in a shared studio for sizing, slant, thickness, and ink color, then copies straight into a document as a floating picture with a transparent background.

Runs entirely on your own machine from a single HTML file, as a browser page or a desktop app window. Nothing you make ever leaves the device, which is the whole point for something as personal as a signature. Free for members to use; ask an officer for a copy and the walkthrough.

Security tooling Running

AEGIS · DeepRoot

Endpoint telemetry with an AI triage layer. Installs and manages Sysmon on a Windows host, ingests events in real time, assigns severity by rule, then escalates the interesting ones through a tiered model chain rather than paying for the whole firehose.

Everything lands in a dashboard over a REST and WebSocket API. Storage is tiered so only the events that matter are kept.

Hardware · Digital Orukami Porting

ECHO_MiniDaemon

A 4 inch desk screen that shows how much Claude Code you have left, with a creature of its own. Created and designed by Digital Orukami on the Waveshare ESP32-S3-Touch-LCD-4, inspired by Clawdmeter: the ECHO creature and its animations, the approve from device flow and the hub pages are original work, fed once a minute by a small Windows tray app over Bluetooth. No account scraping, no cloud in between.

The port is one folder of board files behind the upstream hardware layer: an RGB parallel panel scanned from PSRAM, both expander revisions detected at boot, one button doing the work of three. Panel and Bluetooth run today; touch and battery are pending, and the page says so.

Agent architecture · Digital Orukami Running

Fox (PUMBA)

Since I started messing around with local models, the ideas have not stopped. The joke is that the breakthrough usually lands around 12:30 AM, and the digital caffeine high that follows keeps me up until 3 AM with class at 9:30. Fox (PUMBA) is the project that grew out of that: it started as a tiny chat agent I ran off a cheap GPU from a local garage sale, proof to myself that a local model did not need to feed the corporate expansion of AI consumerism to be useful. It grew from there, more data, then the entire ECHO landing page and club history, then a second GPU so it could toggle between two brains instead of one. This past weekend the v2 build of Fox got upgraded, which is why I want to give the project some context before getting into the new hardware for v3.

Fox is a local inference AI agent I design and build as the ECHO Club's guide, running fully on self hosted hardware so nothing depends on a cloud API. The core is a dependency light Python relay that serves a local Ollama model over a loopback endpoint with a strict honesty doctrine: when no model is reachable it says so plainly and never mocks a response. It offers a two brain toggle, a fast 3B model for quick answers and a deeper 8B for harder ones, retrieval over a curated club knowledge base via tool calling, an append only message log, and a custom HUD style comms dashboard with a compose stage, live status, and a transcript record. The agent has an identity of its own; in the hero screenshot it chooses its own nickname, Fox. The current work is a voice capability: not off the shelf text to speech but a composed, expressive voice language built from a small alphabet of sounds, tuned on an inflection bench and streamed live. The project spans agent architecture, local model deployment, retrieval, and voice synthesis.

This past weekend the rig behind Fox took a real step forward. The target is a workstation the team nicknamed unitzero, built around a Threadripper TRX50 platform running Linux, aiming for 72 GB of pooled VRAM across three Intel Arc Pro B60 GPUs. Two of the three cards are in and already pooling memory across PCIe to run a 32 billion parameter Qwen3 model, quantized, that will not fit on either card alone: a single card load fails and the two card pooled load succeeds. On this prototype path, llama.cpp's Vulkan backend, generation runs around 10 to 11 tokens per second, with a faster Ollama compatible serving stack planned but not yet installed. Fox's relay only speaks Ollama's API shape, so rather than rewrite Fox I built a small translator that converts the rig's API shape into the one Fox already expects, and I proved it through the live dashboard itself: a real message sent through Fox's UI got a real reply back from the new rig. The honest, charming part is that when I asked Fox about its new hardware, it said it did not know, because its own persona files still describe the old hardware and only the model backend had changed, the honesty doctrine working exactly as intended. The third card, a faster production grade serving stack, and tensor parallel validation across all three cards are next. This is real and working, not finished.

Language design · Digital Orukami Live

Fox Script

A written alphabet for an AI agent. Fox, the club's agent, already shows what it is doing through eleven hexagon letters that move across its display. Fox Script is the other half: a full A to Z plus digits that stays still and can be read at leisure, built on the same hexagon lattice so the two read as one language.

Three competing glyph systems, each generated from a single rule, benched side by side with every number measured rather than eyeballed: how far apart the closest two letters sit, how small each system stays readable, and whether words keep their gaps. Every sign on the page is computed at load, and a lite version is one click away.

For SGA · Student body initiative Collecting

Campus Food Access Survey

A survey built to survive a hostile read. SGA is making the case for keeping the Peak Cafe open during the summer semester. ECHO built the instrument, the page, the anonymous intake and the public results board.

Two tiers: four questions anyone finishes at a poster, and a full set for those who choose to continue. Everything that could be waved away is measured instead, the students who don't use the cafe, what a fee would actually cost per credit, whether work study awards are going unplaced in the semester it is shut. Results publish whatever they show, including the findings that argue against it.

Wireless survey Running

AUSPEX

Situational display for the airwaves. A hex mesh radar and live compass over Kismet, turning a wireless survey into something you can actually read at a glance instead of a scrolling log.

Built as its own application alongside AEGIS, sharing the same design language and the same relay fleet underneath.

Hardware · Digital Orukami Running

ECHO_PEX

A pocket WiFi scan and sweep detector that fits in a hand. Built on a LilyGo T-QT Pro: an ESP32-S3 radio sits in promiscuous mode, hops the 2.4 GHz channels, and watches management frames for the shape of someone scanning or flooding the air around it, then flags it live over serial to a read only LAN dashboard anyone on the network can open for free.

The dashboard now carries its own capability catalog: forty entries, each a card you open, covering everything it already sees and everything the same hardware could still be made to see. It also carries a small watcher that notices the board the moment it is plugged in, launches the dashboard itself, and opens the browser without anyone touching a terminal. Sibling project to ECHO_LabDaemon, same house, different board. Private repo for now.

Agent experiment In progress

TokenBurner

Several AI agents sharing one Slack workspace. One app, many agents, routed by who you mention, so a channel can hold a conversation with more than one of them at once without each needing its own install.

It started as internship work and is being rebuilt as something students can follow along with, and run themselves. The interesting part is not the bot; it is what happens when several agents share a channel, how you route a message, how you keep them from talking over each other, and what you let each one actually do.

This site Live

ECHO Club

The club's own front door. A static site with a Google Apps Script membership gate behind it, sign in, roster, one click approval, and the member deck.

Built on shared engines rather than per page code: one page registry drives the nav, the menu and the member gate; one engine drives every detail card; one builds the section rail from a page's own sections.