The words for what they love.
Millions of people speak through AAC: tablets and picture boards that turn a tap into a voice. Those boards are built for needs: eat, drink, toilet, pain. Fringe is an agent that gives them the words for Bluey, Sinatra and last night’s game, so they can do what people mostly talk for: connect.
Live agent · 10 Qloo capabilities · exports to Open Board Format · no account, nothing stored
Americans can't reliably meet daily communication needs with their own speech.
Beukelman & Mirenda, via the National Academies (2017)
reason adults communicate is social closeness, ahead of information, etiquette or needs.
Light's purposes of communicative interaction
of autistic children remain minimally verbal, even after years of support. Many never get a board that knows what they love.
Tager-Flusberg & Kasari, Autism Research, 2013
systems and poor fit are among the factors 275 AAC specialists most linked to people abandoning AAC.
Johnson, Inglebret, Jones & Ray, AAC 22(2), 2006
norms shape most AAC practice; therapists have asked for country-specific tiles since 2022.
Kulkarni & Parmar, AAC 33, 2017 · cboard#1267
And a board doesn’t stop speech: across 27 cases of AAC intervention, none spoke less and most spoke more (Millar, Light & Schlosser, JSLHR, 2006). Today, fringe vocabulary is still added by hand, one button at a time.
A board without culture is technically functional. A board that knows what someone loves is how they make friends.
Try it with a real-feeling person
Each board below is built live by the agent from a one-line note. Tap a picture to hear it speak; tap i to see the Qloo evidence.
Her board says “toilet” and “juice”. It can’t say “Bluey”.
Loves Bluey, Taylor Swift, Encanto, Minecraft, mac and cheese, swimming.
Build Maya’s boardHis classmates talk about the game last night. He has no button for it.
Loves Spider-Man: Across the Spider-Verse, Travis Scott, NBA 2K, Rocket League, deep dish pizza.
Build Leo’s boardHis board came from an American template. It doesn’t know Chhota Bheem.
Loves Shah Rukh Khan, Chhota Bheem, cricket, Pokemon, Arijit Singh, butter chicken.
Build Aarav’s boardForty years of Sinatra and Christie, and a board full of “pain 1–10”.
Loves Frank Sinatra, Casablanca, The Great British Bake Off, Agatha Christie, clam chowder.
Build Ruth’s boardAn agent, not a prompt
The language model handles language. Qloo decides what goes on the board. Every step is streamed to the studio, so you can watch it work.
- 01
Understand
Reads one sentence from a parent or SLP and pulls out the shows, songs, games, books, foods and places.
Groq · gpt-oss-120b
- 02
Resolve
Finds each favourite in Qloo's 250M-entity taste graph, with a real picture for the button.
GET /search
- 03
Expand
Follows taste across domains: Bluey → Charlie and Lola, Frozen II, Olivia Rodrigo. Explainability says which favourite led to each pick. Age-rated for kids.
GET /v2/insights × 6 · explainability
- 04
Localise
A heatmap of where their kind of taste lives in their city, then parks, pools and museums inside those hot-spots, snacks they'd ask for, and where to eat their favourite food.
urn:heatmap · geohash · brand · tags
- 05
Read the room
Checks which age groups each topic lands with, so the board knows Bluey is a parent topic and Sinatra a grandkid one.
urn:demographics
- 06
Bridge
For a named partner, compares the two taste profiles and keeps topics both generations score positively and that draw on both people's favourites.
compare · audiences · explainability
- 07
Verify
The LLM can only phrase and choose from ids Qloo returned. Anything else is dropped before it reaches the board.
closed-set check
- 08
Keep it current
“She’s into dinosaurs now.” A tool-calling agent turns that into Qloo lookups. Or let her choose: This or that shows two Qloo pictures and follows what she reaches for.
Groq tools · choices as signal
Why it needs a taste graph
Ask a language model alone and you get a reasonable list of words. The studio runs that experiment live, then checks every suggestion against Qloo. The difference is everything an AAC button actually needs.
See the side-by-sideBuilt the way AAC is practised
Core words never move
24 high-frequency words, each with a symbol, in fixed positions so motor plans stay stable. Help, stop and all done are always there.
A human approves
Favourites you named go straight in. Everything the agent suggests waits for a parent or SLP to keep or remove.
No lock-in
Exports to Open Board Format (.obz), the open interchange format for AAC boards, with pictures and phrases.
Talk with someone
Build a page for one conversation partner: shared themes, joint topics, and questions to ask them.
Let them choose
This or that: two pictures at a time from Qloo, the next pair following what they reached for. For people who can't tell you what they like.
Their family's language
Hindi or Spanish core words and phrases, spoken with a matching voice, using Qloo's localised names.
Learns from use
Taps are counted on the device. "More like what she taps" sends the most-used buttons to Qloo as the signal.
Evidence on every button
Affinity, which favourites led to it, how widely known it is, age groups and address. Nothing is unexplained.
Nothing stored
No account. Boards live in your browser; the server keeps no personal data.
Who it’s for, and where it goes
AAC hardware and apps are routinely funded by insurance and schools. The labour of making them personal is not: it falls on parents and on speech-language pathologists with full caseloads. Fringe turns that labour into a one-sentence review task.
Families
Free in the browser. Describe your person, review the suggestions, export to the AAC app you already use.
SLPs & schools
A caseload studio: one board per student, refreshed each term as tastes change, with partner pages for classmates and aides.
AAC platforms
Already an MCP server any agent can call (build_aac_board, talk_with_partner), and boards export to Open Board Format: the taste layer AAC never had.