THE BENCH

Browdini

A wooden magician automaton that does card magic for a phone camera, with no hand in the scene

A short, big-headed showman on a tall stage table, like a ventriloquist dummy that fell off a workbench, built to do one card trick for a phone at audience height. Six cards lie fanned in front of him. He reads the sign, looks the cards over with two eyebrows doing all the acting, dips his head, both mitten hands sweep in and cover the fan, the hands lift, and a different five cards are lying there: the one you thought of is gone. He gloats, notices something's wrong, throws a tantrum, and a burst of cards sprays out of his mouth onto the table. Twenty-seven seconds, kazoo and slide whistle, never a word. An ESP32 and ten hobby servos run it; the body is 3 mm Baltic birch cut on my laser and skinned in kraft paper, with washers for eyes.

The secret is under the felt

The whole front of the table top is a 240 mm turntable under black felt. The six-card fan is glued to its front half, the five-card fan to its back half, which hides under a low shelf the magician stands on. While the hands cover the front, one servo turns the disc half a turn. The hands never touch a card. Nothing the camera can see ever moves except a circle seam in black felt, viewed edge on. The tantrum's card spray is a rubber-band tray in the mouth with a servo as the catch, not the spring, so a small servo can hold it and let it go. Two eyebrows on two servos do the face: level, smug, suspicious, furious, astonished; the scowl is the one the trick leans on, so each brow pivots near its outer end and the inner ends drop.

Where it stands

Everything is a file and nothing is a thing yet. The design record, the wiring, every laser sheet and every printed part come out of one parameters file, and the firmware's whole show is a table of seconds and moves that runs on the PC against a stubbed servo board, so I can watch the timeline before any wood exists. A Codex audit on October 23 found four things in the motion (a guard so the swap can never turn while the arms are open, a stop for the mouth tray, the card leaving before the head lifts, a mislabeled channel); all four are in. Next is mine: cut the kerf coupon, cut the five sheets, print the eight parts, wire it, then flash with me watching. Tricks 2 to 4 (a prediction, card to mouth, a vanish) are written and wait on trick 1 working. Lineage is on purpose: it borrows the idea of an automaton magician from Automabot and keeps none of its look.

• ESP32 + 10 servos
• Laser-cut birch + 3D print
• Automaton

No public links yet.

Last activity 2026-10-23

TILT 64

A two-handed tilt game console where the LED panel is the controller

An Adafruit Matrix Portal M4 on the back of a 64 by 32 RGB LED panel, a thumb button for each hand in the side handles, a piezo buzzer, and that's the whole console. You hold the panel upright in both hands and tilt it to play. MARBLE is a marble on a tilting board four screens wide: holes, coins, ice, sand, pink orbs on patrol, checkpoints, five hand-built levels. RAIDER is a side-scrolling shooter in the spirit of Scramble: sky, cave, city, tunnel, rockets that launch as you pass, saucers, and fuel tanks you shoot to keep flying. A setup screen learns how you hold it, so the board can sit at any angle on the back. High scores live in the chip's flash.

How it plays

Turn the panel like a steering wheel and the marble rolls left or right, or the ship flies forward or back. Tip the top edge toward you or away and the marble rolls down or up, or the ship dives or climbs. Left thumb hops the marble over a hole or fires; right thumb brakes or drops a bomb. Hold both for a second and you're back at the menu. The menu also has a SPEED setting, 1 to 5, and a sound switch, both remembered. A DIAG screen shows the raw tilt, both buttons, frames a second and the panel's refresh rate, which is how I prove the wiring.

Nine days in a simulator, then one afternoon on the panel

The first firmware was written on October 13 and lived in a PC simulator for nine days before it touched the panel: a bot plays both games with scripted tilt and writes contact sheets of the frames, and those sheets are the pictures on this card. October 22 was the first flash. I played it, and that one afternoon threw out most of the design. Play moved from lying flat on a table to holding the panel upright. The lunar lander game got retired (my words: that game sucks) and RAIDER was written and flashed to replace it the same day. Everything was far too fast, so every speed got halved and the SPEED setting went into the menu. The menu stopped moving on tilt because a panel lying flat read as a hard tilt and clicked through it nonstop. Four flashes in a day. Once the board dropped into its bootloader mid-setup; a re-flash brought it back, nothing damaged, cause not known.

Where it stands

It runs on the real panel and I've played both games on it. Not done: the ten-step bench checklist (the frame rate on the DIAG screen, each handle button, the dark-scene check, whether it remembers its setup through a power cycle), tuning by play, and whether the buzzer straight off a pad is loud enough. The simulator says 120 frames a second; the panel itself hasn't been measured. More marble levels and a third game are on the list.

• Matrix Portal M4 + 64x32 LED panel
• C++ (PlatformIO)
• Game

No public links yet.

Last activity 2026-10-22

DropSpider

A porch spider that drops on whoever walks up to the door, then climbs back up and waits for the next one

A Halloween prop for the front porch. Somebody walks toward the door, a 24 GHz mmWave human presence radar sees them coming, and a foam spider drops out of the dark -- half a meter to a bit over a meter -- timed to land about when they reach the doorway. Then it winds itself back up and waits for the next one. While it waits it draws no power at all: a small servo drops a finger into a printed ratchet, and that holds the spool. An ESP32 runs it, a small stepper motor does the lifting, and the whole thing hangs on one 6 mm rod. Also, it serves its own web page for setup and takes firmware updates over WiFi, so once it's up on the beam I never have to plug it into a computer.

How it works

Motor, coupler, rod. The spool rides the rod on a one-way bearing -- it grips turning one way and spins free the other. On a drop the motor spins fast in the unwind direction and the spool runs down behind it. The bearing lets the spool fall behind the motor but never get ahead of it, so the motor decides how fast the spider falls and where it stops. On the way back up the bearing grips and the motor winds the spider home, where a small stopper on the line lifts a little hinged flap and clicks a switch. The servo finger drops into the ratchet, the motor shuts off, and it waits. The radar is a DFRobot C4002, a 24 GHz mmWave presence sensor that reports how far away someone is and how fast they're coming. The drop fires only on someone walking toward the door, close enough that the spider lands as they get there. Walking away, walking past, or standing still does nothing.

The part that fought back

Letting go. With the spider hanging, its weight pins a ratchet tooth against the finger, and a little servo can't pull the finger out from under that. So the motor winds the spool up a twelfth of a turn at the same moment the finger swings out, the tooth backs off, and the finger slips free. The other surprise: I planned to let the spool free-fall with the motor switched off, but a stepper that's switched off drags too much. So the motor stays on and leads the drop. And the line: I started with fishing line, and its coil memory sprang loose wraps off the spool. It's bonded nylon upholstery thread now, limp, with about a fifth of stretch in it, and that stretch is the shock absorber when the spider hits the bottom.

Where it stands

It's built and it drops. Since the middle of October it has hung on a mount over the front door inside the house, firing off the radar: one night it dropped 25 times and wound itself home every time. The thing it has to get right every single time is the rewind. If the line misses the home switch the spider hangs part way down, so the last two turns of every rewind now crawl, a missed switch gets one slow retry, and three presses on the switch bring it home by hand. The board is on its tenth firmware flash, all over WiFi; the eleventh (a sleep mode by pressing the switch five times, a heartbeat blink while it sleeps, a tidier page) is built and waiting for me to watch it go on. Still to do: the formal bench tests, then the porch beam and the walk tests before Halloween. On October 14 the repo got its publish pass: MIT license, a one-page build guide, every doc in my own voice.

• ESP32 + stepper + mmWave radar
• 3D printed
• Halloween prop

• Code

Last activity 2026-10-14

Attune

MusicIP Mixer rebuilt: one song in, a mix that sounds like it out, then one click to Plex or the car USB

MusicIP Mixer was the best playlist tool ever made. You handed it one song and it built a mix of tracks that sounded like it, from the audio, not the tags. The company folded around 2010 and I kept it alive for fifteen years anyway. Attune is my replacement. It's a Windows app that reads my library of about 21,000 analyzed tracks and builds a mix the way MusicIP did, with a built-in engine that matched or beat the original in a blind listening test. Then it does what MusicIP never could: when the mix sounds right, one click puts it on Plex as a playlist, or one click copies it to a USB stick for the car or the kids' players. No cloud, no account, one exe in one folder.

What it is

Attune is a desktop music app for Windows that runs from one exe in one folder and never phones home. It started out laid out like MusicIP Mixer and now borrows MusicBee's shape: a navigator tree on the left, genre, artist and album panes, a track table, a right rail with album art and what's up next, and a player bar along the bottom. The point isn't the player, though. The point is the mix. Pick a seed song, the one the mix grows from, and it ranks every track in the library by how close it sounds. The main ingredient is a neural audio fingerprint (how the track sounds, not what it's tagged), then genre, how far apart the years are, and five sensations measured from the sound itself -- Pulse, Glow, Heat, Voice and Grain. Style, variety and artist spacing are the same knobs MusicIP had. On top of that it has a radio mode, saved recipes, a Genius button that picks the seed for you, and mixes that blend two or more seeds.

Why I built it

My car plays mp3s straight off a USB stick, by file and folder. So do my kids' players and the boombox. It's the most dependable music setup I've ever owned, and the missing piece was always the playlist brain. MusicIP was that brain, but getting one mix onto a stick was a chain: export from MusicIP, save the playlist file, import it into MusicBee (another music player), rebuild the playlist there, export the files, copy them over. For every mix, for fifteen years. Attune grew out of the folder where I was trying to get playlists out of MusicIP, became a from-scratch rebuild of that brain checked against my real MusicIP library, and turned into the thing that replaces it. And when the old engine finally won't run anywhere, the built-in one is mine, under an MIT license (free to use and change), and outlives it.

Two engines, and the one that won

Attune comes with two engines and a third you can switch on. If a real MusicIP Mixer is running on the machine, Attune finds it and borrows its engine. If not, it falls back to its own, called V2: a neural network (CLAP) as the ears, a timbre fingerprint alongside it, and light rules for tempo, era and genre. I captured real MusicIP mixes for five seeds and ran a blind listening test across five engines, MusicIP included. V2 matched or beat the original by ear; raw CLAP alone came in weakest because it wanders across tempo and decade. One rule came out of that: taste is judged by ear, never by a score. The third engine was trained to copy MusicIP's own rankings. You can pick it, but it doesn't become the default until it passes the same test.

Where the mix goes

Every list that's a playlist ends in a SEND TO bar, the same spot a Winamp playlist window kept its buttons: how many songs, how long, then two buttons -- Plex playlist and USB / car folder. Press Plex and the playlist shows up on the server; the bar says 'On Plex now' and how many made it. Press USB and it asks which folder (it remembers the last one), then copies the files over, numbered in mix order with names a car stereo can read, with a progress bar right there. Behind More are the extras: save or download a playlist file, and the folder mirror. The folder of mp3s I keep for the car gets matched to the Plex library by file name, then by tags, with the track length as the tiebreaker and never a guess, and it only writes after I've looked at the list. Playlist names can carry the date, so DrivingTunesUSB ({date}) makes a fresh dated playlist each time and leaves the old ones alone.

Show the magic

Since 0.2.1 the mix shows its work. A button called Show the magic draws the whole library as a ring of dots around the seed, the ones that sound enough like it inside a dashed cut-off ring, and beside it a funnel that reads like a sentence: 21,255 in your library, 21,118 can be mixed, 258 looked at, 150 taken, and only 4 sound enough like Tootsie Roll. The cut-off is a slider tied to that ring; drag it and the list re-mixes while you're still holding it. Every pick has a Fit column and a Why this pick line in the recipe's own words. Star ratings now live in the music file itself and reach Plex, and Edit Tags edits every field. The newest thing, in the source since October 6 and not in a release yet: every song's loudness measured the way radio measures it (ffmpeg's broadcast meter, 80 minutes for the whole library), so Radio's hold-steady setting keeps 88 songs in 100 about as loud as the seed, up from 70.

0.2.0: the ears are trained now

Turns out the copy of the CLAP model every version up to 0.1.2 listened with had never been trained. It still made 512 numbers a song and the mixes still came out related, but those numbers weren't what a trained ear hears. 0.2.0 has the model its makers actually trained, and a library scanned with it gets a new recipe, the sound profile: the fingerprint goes through a small extra layer that learned which artists ListenBrainz listeners play together, then genre, year and the five sensations sort the order. Tempo doesn't count any more. I picked it without listening, by scoring over 150 versions on which artists real listeners play together, and it finds music people like together about half again as often as 0.1.2 did (49 picks in 100 against 33). That's a sign, not a listening test, and nobody's judged it by ear yet, me included. A library from 0.1 keeps working the way it was and can be brought across from Preferences, about 1.6 seconds a song, so hours for a big one.

Where it stands

Version 0.2.1, October 5, 2026; it fixed a 0.2.0 whose desktop build could not start. There's an installer and a zip on GitHub for anyone who wants it. The exe runs on its own, no Python install behind it. My library is about 21,000 songs, all fingerprinted again with the trained model, with 21,118 in the mix pool. Every change has to pass the same test: sixteen saved playlists have to come out exactly the same as before, plus 429 automatic checks. What's still owed is a listening session -- the new sound profile, the blend and adventure mixes and the learned engine have been measured but never heard, and that's the only test that counts.

• Windows
• Python
• Music

Grew from MusicIP Playlist Export

• Code
• Download

Last activity 2026-10-06

IRIS

A robot face that follows you, talks back, remembers the house, and puts the music on

Most home assistants are a microphone with a personality bolted on. IRIS is the other way around. She's a face first: two round displays with eyes that find yours and follow you across the room, a mouth that moves with the words, three lights that shift with her mood, and a voice that trades quips with adults and plays I Spy with my kids. Ask her something and she starts answering in about a second and a half instead of waiting on the whole reply. Say stop and she stops. This summer she learned two more things. She remembers what the house tells her, and she'll put on a song from my own music library when you ask for it by name. None of it touches the cloud.

Why she exists

I built voice assistants before this one. They worked. Nobody wanted to talk to them. A voice answering questions from a speaker is a tool; nothing about it feels like anyone is home. Then I put a pair of animated eyes on the desk and gave them a little sensor that finds faces. Eyes that animate are interesting. Eyes that find you and follow you across a room are something else. That was the moment. IRIS grew from the eyes outward: a face, then a voice, then a personality. My day job is mental health, and IRIS borrows from how people actually attune to each other (timing, tone, a face that matches the words) and treats that as the design spec, not a metaphor.

What's inside

A Raspberry Pi 4 runs the show: the Hey IRIS wake word (the phrase that gets her attention, and I trained that model myself), the microphones, the camera, the lights, and a web control panel that has grown to seventeen tabs. A Teensy 4.1, a small microcontroller board, draws the eyes and the mouth, with 25 eye styles to pick from. A second Teensy pans the head. The heavy thinking (turning speech into text, the language model that decides what she says, and the voice) runs on GandalfAI, the workstation on the home network. No cloud, no subscription. Every reply starts with one of eleven emotion labels, and that label fires the eyes, the mouth, and the lights at the same instant, so the face means what the voice says. There are two of her, a sharp one for adults and a warm one for kids. The kids' one plays I Spy and Guess My Drawing with the camera, and Rock Paper Scissors by reading your hand with the same sensor that tracks your face.

She remembers now

Since late July she keeps a memory of what the house tells her. Ask what we talked about last weekend and she answers from the real record. The first time it worked in the room, on July 30, she named a family wedding we'd been to and the two buttons somebody found on the floor, then stopped. When her record comes up empty she says so instead of making something up. Things you tell her on purpose, a remember that said to her face, get quoted back word for word. Things she merely overheard only color her reply. Say forget that and she takes a fact back. She understands three days ago and the other day. Also, the control panel has a Memory tab where I can read what she holds, forget any of it, or promote it, which means treat something she overheard as if I'd told her on purpose.

Play Free Fallin by Tom Petty

Ask her to play a song, an album or an artist and it comes out of a separate speaker, pulled from my own music library on Plex (the app that serves my music at home), about 21,000 tracks. The first build, in August, pushed music to a Bluetooth speaker. A day later I moved it onto a small network speaker board (an ESP32) through Music Assistant, an add-on for the Home Assistant home automation app, and Bluetooth came out of the path. She looks the request up in the library, checks the artist, and only says Playing X after the player reports the stream is actually rolling. If the library doesn't have it she says so rather than playing something close. Play X radio queues up similar tracks. Pause, next, stop and volume work by voice, a plain stop on its own silences everything, and if nobody has touched it for three hours she turns it off herself.

She waves back now

September was about what she does when nobody has said a word. Wave at her and she says hello back, in one of ten languages, most of the time; the bank is twelve short greetings I picked by ear after the synthesized ones and then the real recorded ones both got thrown out. Hold up an open palm and she hushes: she cuts whatever she's saying and holds every unasked line for 45 minutes, with the put-out face and a magenta blink instead of a word. Say hush or pipe down and the same thing happens. The wave hello runs only while the greet switch on the control panel is on, and it was off the last time anyone looked. Then the light seam: a strip of 39 LEDs along the front face that mirrors her three small lights and blooms when someone arrives, so she can show she noticed you without speaking, and sleeps as a starry nightlight. The printed part came out of parametric CAD, the strip is wired to the eye board, the Pi drives it, and what each cue actually looks like on the strip is still to be checked by eye.

Where it stands

The session counter is past 306, and the changelog carries 241 of those numbers under their own headings. The code on GitHub is the real running source, not a demo. Her personality gets tested too: a small test kit sends a fixed set of prompts to the language model and marks every answer that drops the emotion tag, says as an AI, runs past 80 words, slips into therapist language, or swears. The recent work has been the unglamorous kind: a USB fault took her face offline for most of a day at the end of August and nothing noticed, so there is now a watchdog (a small program that checks she's still there and resets the USB connection when she isn't) and a full copy of her memory card on a monthly schedule, first automatic run due in October. The demo video is still owed. A few music checks still need my ears. The eye tracking sensor she started with was discontinued, so its replacement was built and proven in its own small project, CyclopsGaze, and then swapped in.

• Raspberry Pi 4 + Teensy 4.1
• Python + C++
• Robot

Grew from TeensyEyes, MusicIP, whose matcher code the music feature borrows Led to CyclopsGaze Runs on GandalfAI

• Code
• YouTube

Last activity 2026-10-03

The Schmidt Family Constellation

Five hundred years of one family in one file: map, timeline, tree, every claim sourced

Five hundred years of my family, pulled out of one Ancestry tree and a box of scanned papers, and turned into a single web page you move through by map, by time and by text. 646 people, 196 places in ten countries, 13 ocean crossings drawn as arcs, and every sentence tagged with what it rests on. The whole thing is one file that works with no internet at all, built by a chain of Python scripts that won't publish anything their own checks haven't blessed. It lives at theschmidtwithin.com/family/ behind a passphrase, because the family copy carries real names.

What it is

The page is called The Schmidt Family Constellation. Open it and you get a tree you can drag and zoom, drawn the way Ancestry draws one, plus Stories, Places, Everyone, and an Everything else drawer with eighteen more views: how two people connect, stand in a year, a timeline, what the dates say, what the names mean, open questions. The Places view is a map with every place the record ties somebody to, and the ocean crossings drawn as arcs, 1839 to 1924, twelve of them from the old country to the new and one visit back. A second page, From Schleben to Alaska, tells the story in prose with period photographs. Each page is one HTML file with the pictures and fonts baked in, and it never reaches out to the internet, so a copy on a USB stick works forever.

Why it's built this way

I edit the tree in exactly one place, Ancestry. RootsMagic, a family tree program on my PC, pulls it down, and my scripts read that one database and nothing else. Records travel one way, down. What the archive shows, which picture is the portrait, who stays out of the way on a crowded page, lives in files here, so I never touch Ancestry to change how a page looks. Where two papers disagree, both values stay on the page with their sources, and nothing gets averaged or dropped. When the archive works something out from the dates, it says so and shows the dates it read. The calls about what's true stay mine; the code only decides how to draw it.

What's inside

Everything runs by double-clicking a batch file. update-from-rootsmagic.bat is the one road for a pull (a fresh copy of the tree brought down from Ancestry): it proves the pull came back whole against the last one I kept, rebuilds three editions, shows me what moved, runs the check battery (the tests that re-check every promise the page makes), and asks one question, keep and publish. publish-to-nas.bat won't put up a byte the battery hasn't stamped. save-design.bat carries a color change from the page to the live site in about a minute. Since September 3 I've been able to fix a sentence of family memory on the website itself, from my phone, and pull-live-edits.bat brings those words back into the repo, the project's folder on my PC. 156 transcripts of scanned papers sit in the repo, and the facts pulled out of 81 of them land back on the right people, by name, on every pull.

The numbers

635 people, 583 of them gone and 52 living. 186 couples. 1,888 facts pulled out of the transcribed papers, 2,018 events and 2,552 citations out of the database, 158 pictures and documents attached. Birth years start in 1523. 196 places with coordinates across ten countries, 140 of them in the United States, 23 in Germany, 8 in Austria, and 13 ocean crossings drawn as arcs. 11 chapters told entirely in evidence, 80 surnames explained, 63 papers cited on people's pages. The archive works out 91 people who crossed an ocean, 24 who died young and 112 who outlived a husband or wife, three of them twice, and marks each one as worked out rather than recorded. 562 commits (saved changes to the code) in a month, 13 pulls kept, 39 publishes.

Where it stands

Live at theschmidtwithin.com/family/, which answers with a passphrase box. Because the website is locked, it carries the family edition minus a short withheld list; the page is scrambled on my PC before it leaves, and the NAS in the basement never holds anything readable. A share edition with living people stripped out of the data itself is the only copy that goes anywhere unlocked. Relatives get a texted link that skips the passphrase and asks for their name instead. The tree still grows: a pull kept on September 3 brought 37 more people and put a dot on Liverpool. Nothing about anyone living leaves the family.

• Synology NAS
• Python
• Family archive

Led to the publisher this site uses Runs on The Schmidt Within

• Live

Last activity 2026-09-11

ChronoBloom

A three-ring LED wall clock that nudges you to look up

Three rings of LEDs, 97 pixels, stacked one inside the other behind a printed diffuser. The outer 60 carries the seconds and the minutes, the same track a real clock gives both hands. The 24 and the 12 carry the hour as one spoke, and a single pixel in the middle is the stamen. It runs on a XIAO ESP32-C3 the size of a postage stamp, sets its own time off the internet, dims itself to the room, and you run the whole thing from a web page the clock serves itself. No app, no account, no cloud. Flash it from a browser, update it from a browser.

Why I built it

I lose hours in the lair. I go down to print a coloring page, find a project that needs me, and come back up empty-handed to a six-year-old asking where her pages are. Alarms don't fix that. I swipe them away mid-thought and forget. A slow change of light at the edge of my vision does work, so I built it into the thing I already look at. Focus Reminders are gentle light animations at whatever interval you set. No sound, nothing to swipe away. An alarm demands. A nudge asks. Some days it wins, some days the hyperfocus does. The project will still be there tomorrow. My kids being this age won't.

The part I'm proudest of

The colors. Six of the seven faces are named for real flowers (Moonflower, Cherry Blossom, Ember Dahlia, Lotus Pond, Sunflower, Bird of Paradise) and built from that flower's own colors, tuned on the actual LEDs until they read right. The rings map onto a flower too: petals on the outer ring, throat on the inner, stamen at the center, so a chime or a nudge reads as the clock blooming out and settling back instead of a strip cycling colors. Keeping those colors true while the light sensor dims the clock all day took real math. Most projects just turn the strip down and watch the deep reds slide to orange. This one corrects every color once and scales the frame, so a theme still looks like itself in a dark room.

The small stuff that matters

A per-frame power budget, so a plain phone charger never browns it out. Chimes on the quarter hours. A captive portal on first boot, so no WiFi password ever lives in the code. Updates over the air, with the inner ring as the progress bar. Two buttons on the back as the fallback for the day the WiFi is down. Also, the photos don't do it justice: these are bright, narrow-band LEDs behind a diffuser, and your eye reads them in a way a camera can't.

Where it came from

Steve Manley built the original ring clock in 2015, and the printed reflector that shapes the light into pointed hands is a remix of his design. Then a decade of my own rebuilds: an Arduino Nano edition with a clock chip and two fiddly buttons, an ESP8266 port that ran on the wall for years, an 8-inch build sized until the off-the-shelf rings seat exactly, a 15-inch one scaled to match strip LEDs, and finally ChronoBloom, the ground-up rewrite with the web UI and the focus nudges. The front diffuser is mine. The credit chain is in the repo.

• XIAO ESP32-C3
• C++
• LED clock

Grew from Steve Manley's 2015 NeoPixel ring clock

• Code
• Hackaday
• Video
• Tour

Last activity 2026-08-08

GandalfAI

A Dell Precision 5820 gutted and rebuilt around a water-cooled RTX 3090

This is the machine IRIS thinks with: the speech-to-text, the language model, and the voice all run here, so the robot in the kitchen answers in about a second and a half without a word leaving the house. It also listened to 21,000 songs for Attune. I bought a used Dell Precision 5820 workstation and tore out most of what Dell put in it (the drive cages, the shroud, the stock cooling) to make room for a water-cooled RTX 3090, 64 GB of system memory, radiators, a pump and a reservoir, and lighting. Then I cut and printed the parts to hold it all together.

What went into it

The fan panel started as a cardboard template, got hole-sawed out of acrylic, and took a 360 radiator with three fans on the outside of the case. The curved shroud pieces are 3D printed. The bracket that holds the plumbing is purple acrylic. The light bar is a wave cut from clear acrylic with an LED strip glued along the bottom edge, so the whole thing glows blue in the dark. A lot of the build was just fitting, sleeving, and tubing, one evening at a time, on a towel on the floor.

Why

I wanted a box that could run a big language model, speech to text, and a real voice all at once, at home, with nothing leaving the house. A 24 GB card is the price of admission for that. I built this right before the memory market went sideways, and the whole machine came in at about half what that one card costs on its own today. It sleeps when nobody needs it and wakes on a network packet the moment IRIS hears her name.

• Dell Precision 5820 + RTX 3090
• Water cooling
• AI workstation

Runs IRIS, Attune

No public links yet.

Last activity 2025-03-03