.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
The Lantern: Multilingual Local Voice Waiter
These days, restaurants lose money at the order. Errors cost an estimated 1–5% of revenue [1][2], and 57% of allergy-affected diners skip a restaurant when allergen info is unclear [3]. 62% of operators still call recruiting a significant challenge [4]. Today's fixes each break: menu-tap ordering makes every guest tap every dish; drive-thru voice AI costs $500–2,200 per location monthly [5] and suits one speaker, not a table. In our benchmark, cloud voice agents took 19–86 s per turn, and a model choosing its own steps was right 0 of 12 times on six-turn orders. That's why we built The Lantern, a real-time voice maître d' for the table. AssemblyAI Universal-3.5 Pro streams each turn, keeps 54 menu names intact, shows words as guests speak, and lets them interrupt. The model talks, the menu decides: it sets prices, flags allergens, and answers in the guest's language. It handles "those two," "instead," unsaid modifiers, and a dish sold out a minute ago. The kitchen sees every turn, gets a ticket only when the order is placed, and can accept, refuse, or offer a substitute, spoken back to the table. The LLM and voice run on site; only transcription audio leaves. On a four-dish, $36.50 order we were right 10/10, up from 0/10, and 5/5 on two other computers. Five AssemblyAI sessions ended with the right basket and no stall, up from 2/5. The first reply starts 1.4 s sooner; the voice begins 2.65 s after the guest stops. Transcription is our only cloud cost: $0.45 per session-hour [6], about $15–22 per 1,000 orders. Small sample; real-table trials come next. Next: tourist restaurants in the US, Bangkok, Tokyo, then chains and hotels. Restaurants pay $29–49 per table monthly, about $580–980 for 20 tables. Then more languages, a faster on-site voice, "the usual," and tickets into Toast, Square, or Clover. [1] Loman.ai '25 [2] QSR Pro '26 [3] Nutritics '26 [4] Restaurant Business '26 [5] Beancount '26 [6] AssemblyAI '26
30 Sep 2026
Develarper: Token-Efficient AI Agent
Develarper is an engineered containerized AI agent constructed for the AMD Developer Hackathon Track 1. It utilizes a highly optimized four-layer hybrid routing core to minimize total remote Fireworks API token consumption while clearing the required accuracy threshold. The agent dynamically routes tasks across eight distinct capability domains using zero-token heuristic overrides, localized semantic caching, and a supervised PyTorch MLP classifier utilizing all-MiniLM-L6-v2 embeddings to achieve 100% routing accuracy. Low-complexity tasks (such as sentiment analysis, NER, short summarization, and factual questions) are processed entirely local within the strict 4 GB RAM sandbox using an embedded Qwen2.5-3B model via llama-cpp-python. The architecture implements a "local-first" code execution path that tests code generation and debugging locally, using AST parsing validation before failing back to remote infrastructure. High-order deduction, mathematics, and long-context summaries are dynamically escalated to specialized remote models like kimi-k2p7-code and minimax-m3. Token trimming algorithms are applied globally to strip conversational filler and strictly compress input and output footprints, ensuring maximum token efficiency and robust leaderboard positioning.
13 Jul 2026
