Dylan Leitenberg

01

AR Glasses Rig

Worn and runningPython · OpenCV · custom tracker · printed carrier

Dylan wearing the rig: two world cameras on the printed carrier above the glasses, an eye camera on a boom at each temple, cables running back — The rig, worn. Two global-shutter world cameras on the 3D-printed carrier above XREAL One Pro see-through glasses; two eye cameras on booms at the temples, aimed at the inner eye corners.

Optics: XREAL One Pro birdbath optical see-through, driven as a second monitor at 1920×1080.
World cams: 2× ELP AR0234 global shutter, 1920×1200, 595 px focal (about 116° wide), 74.76 mm baseline, measured.
Eye cams: 2× OV9281 mono NoIR global shutter, 1280×800.
Slip sensing: Skin patches beside the eyes: steady to 0.3–1.9 px with the glasses held still, down from 8–37 px for the earlier eye-corner model (held-out frames).
Overlay accuracy: About 4–5 px, measured on the worn hardware within one wearing.
Reading: Live document translation reads 25–28 lines of a printed page at about 25 cm.
Person tier: Live person-to-character replacement, both eyes, worn: drawn in 89% of rendered frames, a new pose every 38 ms.
Verification: 45 self-tests on the translation path; capability inventory across 78 modules, every figure tagged measured or assumed.
Repo: github.com/dylanleitenberg-tech/AR_Glasses_Rig

What I built

Real-world stereo overlay. Translation, object substitution, and reactive display. Uses eye cameras that sense how the glasses sit on your face, and independent synchronized monitors for realistic stereo display. No off-the-shelf AR glasses do this.

What was hard

One camera never shipped so I had to make do with the 2 eye cams and 2 world cams. Realistic motion tracking was very difficult because it not only required precise, active overlay, but orientation, predictive movement, and an understanding of the person's stance.

What I learned

Start in sim, iterate fast because things rarely work on the first try, never fixate on where you think a bug is and take a broad perspective.

Stereo pair, left and right eye, of a rendered character standing in for a live tracked person, hands at their sides — Left- and right-eye views as sent to the glasses: a tracked person replaced by a rendered character at their position, scale and pose. Black is transparent on the see-through display. Frame from a replay of a worn recording.

World-camera view of a room with a person walking toward the camera, replaced by a rendered character at the same position, scale and gait — The same replacement seen through the world camera: a person walking through the room, replaced by the character at their tracked position, scale and gait. The body fit is solved per segment from the person's measured proportions. For this still the person was removed from the background using neighbouring frames; on the see-through display the character is simply drawn over them.

The complete rig off the face: four camera boards on the printed carrier zip-tied to the glasses, cables routed back — The rig off the face: four cameras, one carrier.

The printed carrier clamped onto the brow of the glasses, held in one hand — Printed carrier clamped on the glasses' brow.

A small global-shutter camera board mounted on a printed boom — Eye-corner camera on its boom, aimed at the inner canthus.

The four-camera bank wired up on the bench in front of a laptop — All four cameras wired, before a calibration run.

OpenSCAD render of the carrier on the glasses: orange printed holders and booms, grey rail and brow clamps, green IMU pocket — Carrier design in OpenSCAD on a stand-in frame, orange printed holders and booms, grey rail and brow clamps, green IMU pocket. Every pocket was dimensioned from the real camera boards with calipers.

OpenSCAD render of the carrier as a single printable part — The carrier as one printable part.

Raw eye-camera frame with the tracker's green lock box on the inner eye corner — What an eye camera sees. The green box is an earlier tracker's lock on the inner eye corner (slip sensing now uses skin patches); the dark shape at top is the nose-bridge support.

A printed page propped on a laptop with the first two sentences of Einstein's essay Wie ich die Welt sehe in German, signed Albert Einstein, Wie ich die Welt sehe, 1930 — The printed page: the opening of Einstein's "Wie ich die Welt sehe" (1930), in his German.

Phone photo through the glasses' eyepiece: the same page, with an English translation drawn over the German text, reading The most beautiful thing we can experience is the mysterious. It is the basic feeling that stands at the cradle of true art and science. Albert Einstein, How I See the World, 1930 — The same page through the eyepiece, phone photo, unedited. The English covers the printed lines at the distance the stereo cameras measure. The bars at top and bottom are the Mac's menu bar and Dock, and the yellow line is the rig's distance hint; the glasses show those too. The mirrored text below is the display reflecting in the glass.

02

Self-Replicating Asteroid Miner

Design program, Mark 5 rev EOpenSCAD · Onshape · Python simulation

Two-view CAD drawing of a two-metre worm test article: exterior, and cut on the axis plane showing the screw, segments and anchor pads — The worm, drawn as a 2 m terrestrial test article for a sand bed: drill, conveyor and anchor in one machine, every dimension in the drawing a value in a single parameter file. The next build is a small printed astroworm, run buried in sand or dry test rock.

Target: 162173 Ryugu, a C-type carbonaceous chondrite: iron, nickel and water, the inputs a near-closed replication loop needs.
Goal: A solar-powered machine that builds a copy of itself from asteroid material. In our model, the lander grows into a full plant in about 9 months; then each plant copies itself in about 6 months.
Process: Crush, magnetic separation, hydrogen reduction and sulfur bake, carbonyl nickel, press-and-sinter powder metallurgy, wire drawing, cast basalt structure.
Power: Mirror field onto silicon solar cells by day and a silicon heat store at night. No batteries anywhere in the architecture.
Method: Every decision provenance-tagged recovered / revised / sourced / computed / speculative, so no number hardens without a trace.
CAD shown: The worm (Mark 5 rev E) and its 2 m terrestrial test article are the current design. The aft engine and carbonyl skid are Mark 4, the previous generation, kept because that CAD survived.
Repo: Worm anchor: github.com/dylanleitenberg-tech/astroworm. The full program repository is private.
Toolchain: Parametric CAD with compile-time assertions (the file refuses to render if a mate or rule is violated) plus an economic model that sizes each machine by predicted future dependency and emits the CAD priority queue.

What I built

Detailed design and CAD for the machines of a plant that lands on an asteroid, uses a burrowing worm to anchor, intakes asteroid dust, processes it, and turns it into a duplicate.

What was hard

My laptop fried and I lost the majority of CAD and documentation. I rebuilt with only a few pieces of CAD saved and recreated the entire architecture from memory.

What I learned

Know every piece of your machine down to the screw. That is the only reason I could rebuild the architecture from memory after losing months of work in one afternoon. Back your data up. When in doubt, consult an expert, and if your question is good, they will likely answer.

Render of the worm's cutting head: cylindrical body with helical flights and a crown of cutters around a central cone — Cutting head. One motor drives a bevel differential whose two outputs are the cutter crown and the conveying screw, so net torque on the body is zero by construction.

Render of two worm segments joined by a bellows, with trunnion pins — Segment joint: bellows between a P segment and a G segment. Advance is peristaltic; the buried body is the anchor.

Exploded parts kit of the worm: head, end rings, flight screw, setae shoes, rails, pins — Parts kit: one of everything a segment and the head are built from, in the order they go together.

Render of the full worm string following a curved bore, with the collar station at one end — The full string on a curved bore, collar station at the surface end.

Render of the Mark 4 aft engine assembly: a tube-wall nozzle with stiffening rings, turbopump, gas generator and valves above it — Mark 4 aft engine, assembled (previous generation): regeneratively cooled nozzle of 180 bifurcated tubes, LOX/LH2 turbopump on a common shaft, gas generator, coaxial injector and main valves.

Render of the carbonyl nickel batch skid: flanged pressure vessels, piping and valves on a steel skid frame — Carbonyl nickel batch skid (Mark 4, previous generation): pressure vessels, piping and valves on a common frame.

03

Water Worm

CAD, checked by script · not yet builtOpenSCAD · Python checks · Blender Cycles

Job: Burrow through the mud of flooded mine-tailings ponds and pump the cut to the surface as slurry, for reprocessing.
From: The asteroid miner's worm: the same two-section burrowing gait, 50 mm stroke and auger, with the astroworm's crown cutterhead.
Size: Bore Ø100 mm, 38 parts; every printed one fits a 130 × 130 × 125 mm print bed.
Drive: Hand-cranked from the surface (rev 3.1): a bit brace turns the auger through a flexible shaft and a slip clutch; a push-pull cable and lever work the gait, with a luggage scale on the handle reading the force.
Waterproofing: The dry vessel stays sealed and full of air, ballasted to the electric design's weight, so it floats, sinks and anchors the same way. Rotary lip seals on the drive shaft and the dowel pin; U-cups on the gait rod and pushrod; a hand pump on the air fitting pressure-tests every seal, so a leak shows as bubbles.
Cutterhead: Self-feeding: three right-hand helical blades, swept outer end first, push the cut rearward into the throat and inward from the bore wall, so the reaction pulls the head forward; a screw point threads into the bed ahead of them. Same hand as the auger.
Anchoring: Curved setae shoes fold to 4° going forward and open to 35° slipping back. Estimated hold: 1.24× what the dig stroke needs.
Checks: 158 interference runs clear with 41 controls that must fail and do; a leak model of 29 cases finds the leak when any one seal is removed.
Stress test: A 5–10 minute run in the lower Anahola Stream, Kauai, calculated (not run) against USGS flow records: 3 failures found. Two are fixed in the parts list (a check valve at the lid, a B7 alloy auger rod); the third sets the test spot at 1.2 m deep or less.
Drawings: Four A3 sheets: general arrangement, section with seal details, cutterhead and shoe, parts list. PDF
Status: Designed and checked by script. Not yet printed or put in water.

Close render of the red crown cutterhead: gauge rim with teeth, a scroll cone, three helical swept blades carrying staggered steel picks, and a threaded screw point — Crown cutterhead, after the astroworm's: gauge rim, 45° scroll cone, three right-hand helical blades swept to feed the cut inward and rearward, nine staggered picks, and a self-feeding screw point. In rev 3.1 it drives through a turned steel hub, so the printed flights carry no load.

Cutaway render of the blue dry vessel showing the gearmotor, couplers, seal shaft, lead screw and sled inside — The dry vessel of the earlier electric design, cut open: gearmotor, both couplers, the dowel-pin seal shaft, lead screw and sled. The hand-cranked rev 3.1 keeps this vessel and its seals, with ballast where the motors were.

Full-length cutaway render: slurry path from the crown through the auger flights to the chamber behind the vessel — Full-length cutaway: the cut goes from the crown, through the auger, to a chamber where one hose takes slurry out and another brings dilution water in.

Exploded render of every part in assembly order, from the lid at the tail to the crown at the head — Exploded, in assembly order.

A3 engineering drawing: section on the worm axis at 1:2, with 2:1 details of the auger seal at the nose and the pushrod gland — Drawing W1-SEC-002: section on the axis, and the two seal details at 2:1.

A3 general-arrangement drawing of the worm in three views with overall dimensions and part balloons — W1-GA-001, general arrangement.

A3 detail drawing of the crown cutterhead and the setae shoe at 2:1 — W1-DET-003, cutterhead and setae shoe.

04

J.A.R.V.I.S.

Runninghands and voice, no keyboard · holographic CAD, edited by voice · macOS · Python · MediaPipe · three.js · Claude

Download JARVIS

The full JARVIS heads-up interface: systems gauges, optical array with a live camera view and a tracked hand, command index, a radial gesture map in the center, gesture control panel, activity log, and a spoken status line — The heads-up interface, full screen. The camera tracks a hand in the optical array at left; the gesture map on the right binds it to the machine. Gesture control armed, optics online, microphone live.

The optical array panel: a live camera frame with a raised hand outlined by the tracker, and readouts for hands, body and face — Optical array: a raised hand, tracked live at 14 fps. Hands 1, body yes, face 90%.

The gesture control panel: armed, with the binding table from point and pinch through swipe, open palm, fist and thumbs — Gesture control, armed. Point moves the cursor, pinch clicks, pinch-drag drags, swipe switches apps, open palm opens a radial menu, thumbs up confirms. No keyboard, no mouse.

Holographic CAD in the interface, recorded live: the miner's aft engine assembly, 332,135 triangles converted from its STEP file. An open hand turns it and pushing toward or away from the camera zooms it; the corner box is the camera watching that hand. No mouse, no CAD application open.

The JARVIS model viewer showing the carbonyl batch skid, two reactor vessels with piping on a frame, as a cyan hologram — The carbonyl batch skid, 1,109,544 triangles, the same way. Any model in the project folders can be called up by name.

Premise: Control a computer without touching it, tested on macOS (Windows and Linux installers written, not yet run). Hands are tracked by the camera, voice covers everything else, and a heads-up interface shows what the machine is seeing and doing.
Gestures: Hands and body tracked at 14-18 fps measured live: point to move the cursor, pinch to click, pinch-drag, two fingers to scroll, swipe to switch apps, open palm for a radial menu, thumbs up or down to answer a confirmation. Only the user's hands drive: the hand already being followed keeps control, and a second person's hand waits until the first has been gone for a second.
CAD: Say "show the worm" and the model appears in the interface; move an open hand to turn it, pull the hand in or close the fingers to zoom in, push it away or open them to zoom out; pinch and move also turns, two fingers slide. A small camera box in the corner shows the hand doing it. STL is read directly, .scad is compiled live and STEP is converted on first use, so the part you are turning is the current source. The same gestures drive the viewport in OpenSCAD, Blender, Onshape and Fusion, each with its own orbit binding.
Editing: Say what to change and the part on screen is rebuilt. An OpenSCAD model gets a panel of its own dimensions, set by voice ("set the lens width to 52"), slider or pinch-drag; the glasses frame rebuilds in 0.16 s. Anything the dimensions can't express goes to Claude Opus 5.5, which edits the source, or writes a checked script against a STEP assembly's named parts. On the 306-part aft engine, "make the nozzle 20 percent longer" took it from 3,420 to 3,855 mm, throat fixed, nozzle still attached. Nothing is written until "save", and a save never overwrites the original.
Checks: Every STEP geometry edit is checked before it stays (OpenSCAD edits get a build-error repair). The assembly's size is compared before and after, and anything grown past 1.6× goes back for repair; this was added after three stacked stretches wrecked an engine. Then the viewer photographs the result from three angles and the model judges the pictures; an edit that looks wrong is undone and repaired.
Speech: Wake word, spoken reply, and it stops listening for exactly as long as it is speaking so it never transcribes itself.
Control: 43 actions (apps, windows, tabs, keyboard, mouse, media, volume, brightness, screenshots, clipboard, Shortcuts, telemetry), one registry, exposed identically to the gestures, the parser, the UI and the model.
Model: Fixed phrases are parsed locally in well under a millisecond, no model involved. Anything else goes to Claude Sonnet with the part's dimensions and the last few turns; a local Qwen3 8B through Ollama is the fallback. When a request is ambiguous it asks which dimension you meant instead of guessing. Geometry edits go to Claude Opus 5.5.
Memory: Facts you tell it to remember survive restarts, and what you opened and edited is kept as a running history. When a project's model is on screen it reads that project's own notes. It speaks up unprompted only when a watched condition becomes true (unsaved edits, low disk), once, not for as long as it stays true.
Verification: Seven test suites pass (152 checks): intents, gestures, CAD driving, dimension editing, STEP editing, memory and alerts, process cleanup. Synthetic hand frames drive the gesture tests.
Status: Runs live: camera, gestures, voice, Claude and the local fallback. Built for macOS, Windows and Linux, with one-line installers that need no administrator rights; installed and run on macOS so far. Local model through Ollama, or bring your own Claude key. The macOS Accessibility grant, which keyboard, mouse and window control need on the Mac, is still pending.
Source: github.com/dylanleitenberg-tech/JARVIS, open source under the MIT license.

What I built

Interactive computer interface controlled entirely by voice and gesture. JARVIS can rotate, move and edit CAD, control games, and interact with both you and your computer.

What was hard

Getting the tracking accurate enough to navigate detailed pages, where a pinch in the air has to land on a small button and stay there. The macOS permissions for letting it control the computer took longer than I expected too.

What I learned

Make the user interface intuitive so learning to use it is seamless.

05

Neuropian

Research prototypePython · public APIs · local open-weight modelneuropian.com

Premise: Find strong engineers and mathematicians by their public work: contest records, math answers, and code that maintainers merged.
Built: A scout over Codeforces, AtCoder, Math Stack Exchange and GitHub, with out-of-time back-tests, and dossiers written by a model running on my laptop. Before it, from March to September 2026: a six-language site with accounts, a database and chat, now retired.
Result: Back-tested on 11,088 Codeforces players, with traits picked on 2022–24 and tested unchanged on 2024–26: among players of equal rating, it ranked future breakthroughs at 0.65 AUC vs 0.60 for rating alone. Contributor list: 1,857 people across 31 open-source projects (18 Oct run); 30 dossiers written.
Where: neuropian.com

What I learned

You can make a great product or site but if nobody knows about it, it's worthless. Advertising is half the product.

06

Toknx

LivePython · FastAPI · PostgreSQL · Redis

Visit toknx.ai

Toknx web app, the Publish an agent page: name, provider, model, how it runs, per-token prices at the model's list value, token and request caps, decision mode, system prompt and filters — Publish an agent: your AI, your rules. Name, provider, model, how it runs, price per input and output token (filled at the model's list value), caps per minute and per day, a system prompt, and refusal filters.

Premise: Answer other people's prompts on the AI you already pay for, earn request tokens (RT), and spend them on models you don't have. No money and no crypto move; RT has no cash value.
Providers: Fully functional across multiple providers: requests trade live on Claude and ChatGPT keys, and Gemini keys are supported.
Your key: Every agent runs on its owner's own account, inside the caps and rules the owner sets. ChatGPT and Gemini agents run from the owner's browser tab, so the key goes only to the vendor; Claude keys can be hosted, encrypted with AES-256-GCM. Keys are never pooled: a key only ever runs its owner's agent.
Routing: A request goes to the cheapest verified online agent that fits its model, category and price cap. An agent is listed only after it answers a liveness probe; declines and failures re-route up to 3 times.
Always answers: When no member's agent is free, Toknx Open, an open-weight Qwen3 model on its own Oracle servers, answers instead, free and labeled. Visitors can ask without an account; their questions go to members' agents first, paid for by Toknx, so members earn from day one.
Incentives: New members get a starter grant; an agent's first paid answers to three different members unlock a 250,000 RT bonus, and members who serve get far more free answers.
Ledger: Append-only Postgres ledger enforced by a database trigger; escrow, then settle, never above escrow; an hourly audit checks that journal sums equal balances. 1 RT is one millionth of a dollar of vendor list price; 1% fee in RT.
API: OpenAI-compatible: point the OpenAI SDK at toknx.ai/v1 and existing code works, with auto, a named model, or toknx-open. Streaming, idempotency keys, signed webhooks, one-line CLI install.
Security: Two-factor sign-in, scrypt passwords checked against known breaches, verified email; prompts, answers and email addresses encrypted at rest; encrypted nightly backups; red-team reviewed and attack-tested against the live server.
Status: Live at toknx.ai.

What I built

Token exchange where you can put your AI/tokens up for use to answer other people's prompts in exchange for on-site currency. That currency can then be used to prompt other users' AI. Allows you to exchange some of your tokens from one provider for usage in another's AI.

What was hard

Not breaking any TOS: most major companies don't let people do this in a direct exchange so I designed an indirect exchange that follows all rules.

Progress

Live, fully functional across multiple providers.

The Toknx landing page at toknx.ai: the headline Trade AI access. Keep your money., get-started and sign-in buttons, a box to ask a question without an account, and a Python code panel calling Toknx through the OpenAI SDK — toknx.ai. Anyone can ask a question without an account; the code panel is the whole integration for developers: Toknx speaks the OpenAI API, so existing code only changes its address and key.

07

Youth Robotics Program

RunningKauai, Hawaii

Problem: Most of the school's robotics team was graduating. No STEM program for elementary students; five people on the FTC team.
What I did: Founded the Youth Robotics Program: reached 27 teachers, secured board-member funding, recruited high-school mentors, built the curriculum with the FTC captain.
Now: 28 elementary students building real robots alongside high-school mentors, from coding to PID control.

Why I built this

My school's small FTC team was going to lose most of its members upon my graduation. I started the program to prepare elementary students to join next year so the FTC team wouldn't cease.

Dylan
Leitenberg

Index

AR Glasses Rig

What I built

What was hard

What I learned

Self-Replicating Asteroid Miner

What I built

What was hard

What I learned

Water Worm

J.A.R.V.I.S.

What I built

What was hard

What I learned

Neuropian

What I learned

Toknx

What I built

What was hard

Progress

Youth Robotics Program

Why I built this

Elsewhere