The Lantern takes a full table’s order by voice, remembers every word, and keeps the kitchen in perfect sync.
Powered by
The Hackathon
A fully online hackathon run by lablab.ai with AssemblyAI. Every project had to be built on AssemblyAI. The Lantern took the Realtime Speech-to-Text path and brought its own model and voice.
Build Window
30
days, September 1 to 30, 2026, fully online.
Challenge Paths
2
Voice Agent API, or Realtime Speech-to-Text API.
Prize Pool
$10K
$5K in cash and $5K in AssemblyAI credits.
Winning Teams
5
Each wins $1K in cash and $1K in API credits.
Judged on application of technology, presentation, business value and originality. [9]
The Problem
Four pressures on every dinner service, and the technical gap that kept voice ordering from solving them.
Staffing
62%
of restaurant operators call recruiting and retaining staff a significant challenge.[1]
Waiters and waitresses turned over at 131% in 2025, so the floor is always training someone new.
The LanternA voice waiter at every table takes the whole order, so staff can serve instead of writing tickets.
Order Accuracy
1 in 10
drive-thru orders was still wrong in 2026.[2]
Average accuracy in the 2026 Drive-Thru Study was 89.3%. Lanes running AI voice ordering reached 90.8%.
The LanternCode checks every change against the live menu, so a dish, modifier or price can’t be invented.
Allergies
33%
of restaurant servers reported receiving any food-allergy training.[3]
Nearly half of fatal food-allergy reactions over 13 years were caused by food from a restaurant or other food-service establishment.
The LanternAllergies are kept for the whole order, and a conflicting dish is refused before it reaches the kitchen.
Languages
1.52B
international tourist arrivals in 2025, up 4% on 2024.[4]
Each is a guest who may not speak the language the menu is written in.
The LanternGuests order in their own language. Dish names and prices stay exactly as written on the menu.
The Technical Gap
0 / 12
six-turn orders completed correctly when a small local model chose its own tools.[5]
A real ordering conversation
The LanternMemory moves into SQLite and code makes every change: 10 / 10 correct with the same small model.
See It Work
The app’s own screens and the waiter’s own voice, with every reply word for word from the live system. Press play with sound on.
Voices are Kokoro-82M: the waiter uses the product’s own voices, guests use other Kokoro voices. Replies are verbatim from the live stack and from the app’s reply renderer.
How It Works
One guest sentence, traced end to end. The same pipeline runs on an edge box in the restaurant or as the hosted demo; only the model and the voice change.
Hear
Universal-3.5 Pro, streaming
16 kHz PCM over one WebSocket. Interim words appear on the table device while the guest is still speaking.
54 menu keyterms
Interpret
Qwen3 4B on 
Ollama · JSON schema, thinking off
3.5 Flash-Lite · JSON schema
One call per turn returns an intent and a reference such as offered_all. The model never picks tools, writes the order or words the reply.
~1.46k tokens, every turn
Decide
WAL mode · deterministic resolver and validator
Resolves “those two” and “instead”, checks dishes, modifiers and allergens, and writes an immutable revision. Memory lives here, not in the prompt.
10 / 10 six-turn orders
Reply
Template renderer · English and Spanish
Names, prices and totals are read from the saved revision, so the waiter can’t misquote the menu.
Verbatim menu
Speak
Kokoro-82M, local · 24 kHz streamed PCM
Sonic-3 · streamed PCM
A one-second spoken acknowledgement covers the wait, and the real reply queues right behind it.
1,752 ms to first reply audio, p50
Sync
WebSocket /ws/ops to the Kitchen Display
Placed orders reach the kitchen; kitchen decisions carry the order’s revision and are spoken back at the table.
50+ automated tests
Universal-3.5 Pro streaming is tuned for a noisy dining room and a menu full of Vietnamese dish names.
“We’ll take those two, please.” End of turn at confidence ≥ 0.4, then the transcript is formatted.
Keyterms prompting
54 terms generated from the live menu, so dish names come through intact.
Bun Bo HueStir-fried Morning GloryPomelo Salad with Shrimpno fish sauceCaramelized Pork in Clay Pot
Tuned turn detection
Short pauses don’t cut a guest off; a finished sentence is answered quickly.
min_turn_silence=400max_turn_silence=1000end_of_turn_confidence_threshold=0.4format_turns=True
Barge-in
SpeechStarted cancels the reply in progress and stops playback on the device, so guests can talk over a long readback.
Resilient connect
A 5 s handshake timeout with an explicit retry budget. Stalled voice sessions went from 3 / 5 to 0 / 5.
Verified Benchmarks
The six-turn scenario: recommend, “those two”, a sold-out dish, “instead”, a side, “place the order”. The correct final order is 4 dishes, $36.50, placed.
Order Accuracy
100%
10 / 10 runs · RTX 3060
Independent reruns: 5 / 5 on an RTX 5060 and 5 / 5 on a MacBook. [5]
Response Latency
2.6 s
1,752 ms · first full reply audio, p50
Five voice sessions on synthetic 16 kHz speech. Acknowledgement clips are excluded from every figure. [5]
1,341 ms
Model intent, p50
down from 2,111 ms
3 / 3
Refresh recovery
orders survive a reload mid-session
~1.46k
Tokens held flat
turn 1 to 50 of a 4,096 window
5 / 5
Voice sessions
through AssemblyAI, end to end
As of Sep 28, 2026
| Criteria | QR / Table Menue.g. Toast Mobile Order & Pay [6] | Drive-Thru Voice AIe.g. Presto Voice [7] | Generic Voice AgentCloud LLM | The Lantern V2Edge deployment |
|---|---|---|---|---|
| Guest experience | Guests browse and tap on a phone; an order can’t be edited once submitted | Built for short drive-thru orders | Natural speech; free-form replies can misquote items or prices | Natural table conversation in the guest’s language, with instant spoken fillers |
| 6-turn order accuracy | Guests enter each item | Not published for multi-turn table orders | 0 / 12 in our V1 tool-calling build | 10 / 10 runs, 5 / 5 voice sessions |
| Response latency | N/A | 3–6 s† | 19–86 s† | ~1.7–2.6 s |
| Cloud cost / 1,000 orders | $0 | $50–$100† | $35–$60† | ≈ $9, speech-to-text only‡ |
| Audio and data | No voice data | Full audio sent to the cloud | Full audio sent to the cloud | Model and order state stay on the edge; only speech goes to AssemblyAI |
Products change; this table reflects public information on Sep 28, 2026. † Team estimates presented in our deck, not vendor-published figures. ‡ AssemblyAI streaming at $0.45 per hour [8], about 75 s of open connection per six-turn order. The hosted demo also calls a cloud model and voice.
B2B Pricing
For tourism dining, restaurant chains, and hotels and resorts.
Edge Kit
One-time setup
Quotedper venue
Table Service
Monthly subscription
$29–$49per active table / month
Example
20 active tables
$580–$980per month, software only
Proposed pricing, benchmarked against Sinqro and CUKCUK table-app pricing on Sep 28, 2026.
The Team
Five builders from Team Develarper: three AI engineers and two designers who presented the project.
Certificate of Completion
Awarded by lablab.ai, part of NativelyAI, for completing and submitting a solution during September 1 to 30, 2026.
Tap the microphone and order the way you would with your favourite waiter.
Accessed Oct 18, 2026, unless noted.
