Build With AI: Basics · Devpost Learn
Photograph your room and paste a list you must learn in order. Loci puts each item on a real object along one route of up to 12 stops, with a strange little scene for each. Then the lights go out. Say or type the list back: code, not an AI, checks every answer, and each right one turns its stop green.
Stop 1 · The walk
The one-tap example: twelve cranial nerves in a student room. A short walk with the lights on, then lights out and the list typed back, one stop at a time.
scripts/record-demo.mjs. The answers are typed by the script, not remembered by a person; one has a typo, one is an accepted answer.The answer checker, on real answers shared/score.ts
typed occulomotor
→ Oculomotor
Right. One slip in a long word.
typed trochlaer
→ Trochlear
Right. Two letters swapped count as one.
typed auditory
→ Vestibulocochlear
Right. Your list accepts it after “ / ”.
typed 1. Optic
→ Optic
Right. Numbering is ignored.
typed Vitamin D
→ Vitamin C
Wrong. Short words must be exact.
said truck lear
→ Trochlear
Right. It sounds alike: both are trklr.
Each line is what the real checker answered, run on these inputs (isRight for typing, spokenMatch for speech).
Stop 2 · Lights out
Three real screens from the same palace. Code never shows you the answer while you recall; a miss stays dark until you retry it.
The view walks the route, one scene per stop. Stop 4, the wall clock: Your wall clock turns into a little truck-lear, honking ‘Trochlear!’ as it circles the room.
The room goes dark. Say or type each item. Five right so far, so five pools of the room are back.
When every stop is green, the whole room comes back. The result shows the first-try score and both times.
Stop 3 · Who decides
Two AI calls suggest objects and scenes. Plain code picks the stops, draws the route and checks every answer, so the score means what it says.
Plain code, never an AI. It forgives slips in long words, keeps short words exact, and takes your own accepted answers after “ / ”.
shared/score.ts · edit distance per word
Your photo is shrunk to 1280 px on your device, then a vision model names objects and boxes each one.
api/anchors.ts · Gemini, DeepSeek last
It keeps distinct, uncrowded objects and finds the shortest route from the leftmost stop to the rightmost.
shared/route.ts · Held-Karp, exact
From object names only, never the photo, with a sound-alike for hard words.
api/scenes.ts · deepseek-flash
In Chrome or Edge, say the list. The same checker reads what the browser heard.
Web Speech API · shared/voice.ts
Palaces and results live in your browser. The server never stores your photo.
IndexedDB · idb-keyval
Stop 4 · The scenes
The example palace in full, word for word. Strange is the point: an odd picture on a thing you see every day is hard to forget.
yellow raincoat
Olfactory sounds like old factory
Your yellow raincoat turns into an old factory, puffing perfume and humming 'Olfactory!'
blue bed
Optic sounds like op-tick
Your blue bed becomes an op-tick, ticking like a giant eye and chirping 'Optic!'
orange cat plush
Oculomotor sounds like a cool motor
Your orange cat plush grows a cool motor, rolling its eyes and purring 'Oculomotor!'
wall clock
Trochlear sounds like truck-lear
Your wall clock turns into a little truck-lear, honking 'Trochlear!' as it circles the room.
cork board
Trigeminal sounds like try-gem-in-all
Your cork board pushes out three gems, chanting 'Trigeminal!' like a try-gem-in-all choir.
yellow desk lamp
Abducens sounds like ab-duck-ins
Your yellow desk lamp sprouts ducklings that do sit-ups, quacking 'Abducens!'
black backpack
Facial
Your black backpack grows a grinning face, stretching its cheeks and laughing 'Facial!'
striped rug
Vestibulocochlear sounds like vestibule-oh-cock-leer
Your striped rug rolls into a vestibule, where a bull and a rooster leer, blaring 'Vestibulocochlear!'
snake plant
Glossopharyngeal sounds like glossy pharaoh in jail
Your snake plant polishes a glossy pharaoh, who gargles 'Glossopharyngeal!' from a tiny jail.
red kettle
Vagus sounds like vague us
Your red kettle whistles vaguely, steaming the words 'Vagus!' at you.
acoustic guitar
Accessory
Your acoustic guitar sprouts extra strings, wearing rings and scarves as it strums 'Accessory!'
globe
Hypoglossal sounds like hippo-glossal
Your globe spins into a hippo with a glossy tongue, licking continents and roaring 'Hypoglossal!'
Objects found by gemini-3.1-flash-lite and scenes written by deepseek-flash, prepared in advance on 2026-10-04. The room is AI-generated.
Stop 5 · Honest by design
Lifted word for word from the plan and the README.
A proof of concept can only show recall in one sitting.
Your photo is sent once to an AI (Google Gemini, or DeepSeek if Gemini is busy) to find objects. The server does not store it.
Checking a whole sentence, spoken or typed, can't be fair or reliable. Short items only.
Stop 6 · Questions
No. AI finds the objects and writes the scenes. A small checker in shared/score.ts decides green or red. “Vitamin D” never counts for “Vitamin C”.
It is shrunk on your device, then sent once to Gemini (or DeepSeek if Gemini is busy) to find objects. The server does not store it. Scenes are written from object names only.
The one-tap example, 12 cranial nerves in an AI-generated student room, was built ahead of time by the same helpers, so it opens with no AI call. Your own photo goes through the real AI; building usually takes 10–30 seconds.
In Chrome and Edge, through the browser's own speech recognition. Everywhere else you type, and typing always works.
3 to 12 items, in order, 4 words or fewer each. Add other accepted answers after “ / ”, like Sodium / Na.
Loci doesn't claim that. It shows recall in one sitting: the first-try score, the learning time and the recall time. Walk your saved palace again before the exam.
Try the 12 cranial nerves in one tap, or bring your own room and your own list.