AI systems and product engineeringDeliver Globally
TBC is a five-person team in the Bay Area and Vancouver that takes AI from demo to production: realtime voice, retrieval and video generation, the product around them, and the GPUs underneath.
Reno, Jason Zhu, Hongming, Jason Jiang and TomMeet the teamClients · Partners · Pedigree
Advised by former VPs and SVPs of KONE and TK Elevator through our own company, Welink Tech, whose Y Combinator application ranked in the top 10%.
01/Team
No account managers and no handoffs. The five people who scope your project are the five who deliver it and keep it running.
AI systems and algorithms
Experience
Education
Product and AI solutions
Experience
Education
Infrastructure and GPUs
Experience
Award
Capital and operations
Experience
Education
Backend and delivery
Experience
Education
02/Work
What we built, who it was for and where it stands today.
For a global elevator leader, owned end to end: requirements, architecture, launch and operations.
Picks up when a passenger is trapped and answers from the emergency procedures. If a server dies mid-call, another takes over.
Answers model, configuration, price and policy questions in seconds, from hundreds of PDFs.
Fault-code triage, procedures and service policy for field technicians.
Turns “make it feel more premium” into a rendering the client can react to, in hours instead of weeks.
StackTwilio, Azure Speech, GPT-4o, LangChain, ChromaDB, AWS ECS
The group behind Save-On-Foods, PriceSmart Foods, Urban Fare and Quality Foods: ERP, intelligent operations and AI applications, plus the system upgrade and data migration for IDCI, its pharmaceutical wholesaler.
Several large system engagements: ERP build-out and customisation, intelligent operations management, and AI applications.
The legacy ERP moved onto a cloud ERP on Azure, with a full MySQL-based data migration and a stable system after go-live.
StackCloud ERP on Azure, MySQL, ERP customisation
An AI receptionist that takes phone orders and reservations in English, Cantonese and Mandarin, and puts every order straight onto the kitchen board.
StackNext.js 15, TypeScript, PostgreSQL with row-level security, Supabase Realtime, Twilio, Stripe
A text-to-film video platform: text-to-video, image-to-video and batch rendering through an API, with characters and style kept consistent from shot to shot. We built the micro AI data center it runs on.
StackNVIDIA B200, RTX PRO 6000 and RTX 5090, Kubernetes, LoRA fine-tuning
Six Next.js apps in one Nx monorepo, from a DEX and perpetual futures with a live order book to a launchpad and a governance market, with wallets across EVM, Solana and Cosmos.
StackNext.js, TypeScript, Wagmi, Viem, tRPC, Apollo GraphQL, Nx
Build multi-agent teams the way you’d staff a company: an org-chart view, an agent-to-agent protocol, three-tier memory, and a marketplace for tools, agents and whole departments.
StackNext.js 15, React Flow, Zustand, Go, Gin, PostgreSQL
A cross-platform app for spending USDT, USDC and USD1 through a Visa debit card, with Apple Pay and Google Wallet, Face ID or passkey sign-in, KYC, and live transaction tracking.
StackExpo, React Native, TypeScript, AWS, WebAuthn
03/Architecture
Scroll or tap through the production architecture: the path an emergency call takes from a stalled car to a spoken answer, and what happens when something fails.
Someone is trapped on floor 12. The call hits Twilio and streams over a WebSocket to a stateless worker. No keypad menu, no queue.
Azure streaming recognition with tuned endpointing turns speech into text while the passenger is still talking. Barge-in means they can interrupt.
The LLM answers from the emergency SOPs via RAG. Every piece of call state lives in Redis under the CallSid, so any worker can take over. A watchdog reassigns a call ten seconds after a heartbeat goes quiet.
Streaming synthesis overlaps with generation. If a stage times out: shorter prompt → cached answer → human operator with full context. The caller never hears silence.
Latency comparison · one conversational turn
0 – 7.5 seconds
That’s how long the caller waits to hear anything at all. Inside an elevator, it’s the difference between nobody is coming and someone is helping.
Figures are typical production ranges, not a single measurement. Full breakdown in the note on latency engineering.
04/Contact
Tell us the use case, where the data lives and where it’s stuck. We’ll tell you whether AI is the right tool, roughly how long it will take and what to do first. If it isn’t worth doing, we’ll say so.
Reno Yao
Jason Zhu
Hongming Wang
Jason Jiang
Tom Lam
