AI meeting intelligence / GovTech + SaaS · 2025 — present
V-Note Suite
Cantonese meeting AI, sold two ways: on-prem for compliance, self-serve SaaS for growth.
ProductPricing
- HK$1M+
- revenue tractions from enterprize and gov client
- ~HK$500–600K
- deals in active delivery
- 4+
- Entities adoptions
01
Overview
- Role
- Co-PIC/PM
- Timeline
- 2025 — present
- Platform
- SaaS + On-prem
- Scope
- Product Management
V-Note is Cantonese-first meeting intelligence — transcription, diarization, glossary correction, AI-written minutes. It ships as two products on purpose:
V-Note
Air-gapped for regulated buyers
On-prem, zero egress — banks, police and government, who can't send audio off-site.
V-Note Live
Self-serve cloud for every end user and SME
Sign up, connect a meeting, pay by usage — no procurement cycle in the way.
Tasks distribution
My works include product shipping, roadmap, platform-tier design, UX/UI, and Frontend.
V-Note Live
206 tickets total
- Adrian (PM)71.6% · ~147
- AI Engineer83.1% · ~171
- AI Intern34.2% · ~70
- Developer 112.8% · ~26
- Developer 21.8% · ~4
V-Note
225 tickets total
- Adrian (PM)32.2% · ~72
- AI Engineer91.6% · ~206
- AI Intern0% · ~0
- Developer 140.1% · ~90
- Developer 20% · ~0
Why need this products?
- The easy touchpoint — clients get meeting AI on sight, no explaining needed.
- Opening move in 4 B2B deals that bridged into MAGIC and Beever Atlas.
- Sits in the sales sweet spot our reference cases already built — Cantonese LLM, live event transcription.
- Tender won the B2B and starting as an on-prem project.
02
Challenges that we face
Four forces converged at once — not a single problem, four.
Diarization Expectations
Clients expect >90% accurate speaker labels, while there are tech challenges industry-wide(e.g. languages, hardware setup, GPU, Models, etc.).
Bespoke v.s. Scaling
Every customised & on-prem deal dilute R&D team resources while labels and cases are needed for fundraising.
Do clients get AI?
Other AI products like agentic LLM training and AI memory are hard to understand. "It transcribes my meeting!" is the door clients tap into discussion first.
Internal product vision
Biz wants grounded AI products that close deals now; R&D wants to build the technical moat. V-Note sits where those two pull apart.
03
Navigate the Lines
Three aspects, three different moves — what each one demanded on the left, what I did about it on the right.
Deterministic
Gov template
Governmental standard — their own template and terminology, no room to deviate.
Exact match
Get the referenced .docx and match its template and formatting 100%.
Gray zone
Scarce R&D
Not a high-volume deal, and not a company key product — R&D attention stays scarce.
Code-level gate
Lift accuracy for low-resource languages and slang with code-level engineering — e.g. a safe_llm_glossary_output() gate.
Probabilistic
Accuracy bar
Client expects >90% on transcription and speaker diarization.
Benchmark proof
Show the evaluation benchmarking across GPU, ASR and diarization models.
Cutting the GPU requirement on-prem
On-prem buyers rarely have GPU capacity to spare. Moving diarization from NeMo (GPU) to Pyannote (CPU) meets that infra bar directly — and the freed GPU headroom lets V-Note Live run faster and cheaper too.
04
On-prem for trust, cloud for growth
Two products, two buyers — never one codebase to bend between them.
V-Note
On-prem
Enterprise sales-led
Air-gapped, zero egress
Licence-key entitlement
Government & regulated buyers
V-Note Live
Cloud SaaS
Product-led growth
Two buyers, two trust models — bending one to fit the other would have compromised both.
Where V-Note came from
A reseller partner wanted to resell into HK telecoms. Team notes, 2026-07-17: "Originally [the partner] wants resell, so we do V-Note." A channel decision, not a technical default.
V-Note Live — self-serve, growth-first
- Bot dispatch into Zoom, Slack, Google Calendar and Outlook — joins the call without a manual link.
- An AI chatbot surface over past meetings, via the MCP agent tools.
- ASR analysis and generated minutes, ready right after the call ends.
V-Note — on-prem, compliance-first
- Customisable meeting templates — title, summary shape and language, set before a recording starts.
- LLM transcription corrections — a rules gate catches bad rewrites before they reach the minutes.
- Distributor-configurable ASR and LLM endpoints — a reseller like Alibaba Cloud can swap in its own models to upsell its own stack.
What's still broken
05
>HK$1.2M sales volume traction
- HK$1.2M
- Quoted Year-1, one gov proposal
- ~HK$500–600K
- A second gov deal
The B2B opportunity
- 1
V-Note is the easy way in
"It transcribes my meeting" is a product a client can buy without a briefing — the lowest-friction door into Votee's AI stack.
- 2
Upsell Beever Atlas through it
Once the V-Note is deploying, Atlas is the next sale rather than a new pitch from cold.
- 3
Then an AI memory centre
Enterprise memory layer across the accounts, existing platforms, calls, and meetings.
- 4
Business insights come out the far end
Tagged, ingested data the client reads as insight.
06
AI workflow at a glance
Requirements got Socratic-interviewed before they got built; the code got written by an agent I steered, diff by diff.
Claude Code — Agentic coding sessions from a spec, via the OMC multi-agent harness — one logged 283 agent messages against 11 user messages, roughly 25:1.
Deep-interview skill — Socratic ambiguity gating before a line of code — drove the Suite roadmap from ~65% to ~10% ambiguity in one pass.
SpecStory — Session capture plus structured handoff notes, so state survives between agent sessions.
Cursor — Directed a Gemini-backed agent to build a full front-end investor-pitch mockup in one sitting, backend detached.
Linear — Roadmap, milestones and 2-week cycles across both editions — every PR links back to a ticket.
Teammates Involved
Agents did the typing. These are the people who set the direction, argued the trade-offs and reviewed what shipped.
Engineering Lead
AI Engineer
Client-Success Lead
Intern (V-Note Live)