Velq
An AI agent system built to bring a scattered sales process under control.
See it live ↗
Context
Eight years in sales left me with one observation: pipelines lose deals to forgotten follow-ups and poor organisation. SMEs want to fix that with AI without having the time or the in-house expertise to get started. Instead of theorising, I built the tool.
Plug in rather than replace
Follow-ups that slip, poorly qualified leads, reporting done by hand. As a HubSpot Key User and the person who ran the migration onto it, I saw from the inside what the tool does not do: its reporting tells you how many deals are open, not where the pipeline is leaking, and even less what to do next.
Building yet another CRM made no sense. I plugged into the one the teams already use and added only what was missing: the leak diagnosis, then the follow-up plan that comes out of it.
Orchestrator and guardrail
I designed Velq, an AI service for SMEs and sales teams, deployed on Cloudflare Workers and Supabase. The core is a multi-agent system: an orchestrator breaks the work down and routes each piece to the agent whose job it is.
The guardrail sits upstream of the model. Velq treats a fact as verified only if it actually read the company's page, and then cites it with its source URL. Otherwise the prompt forbids it from asserting anything and requires every element to be presented as a hypothesis to confirm on the call.
What goes into the prompt
Content pulled from a web page enters framed as data, with an explicit instruction that it is never a set of instructions to follow, and the extracted signals are sanitised before they get there. It is a defence against prompt injection, written because an agent that reads the web also reads whatever was left there for it.
In production
The agents are documented and versioned, the B2B content is published, and the interface is served in French, English and German from velq.fr.
The trade-offs
Orchestrating agents comes down to deciding who handles what, where a human reviews and when an output has to be blocked. The same trade-offs as designing a product.
Highlights
- 17 specialised agents driven by a central orchestrator
- Company research before a call, lead qualification against an ICP, sequence strategy, meeting prep
- A fact counts as verified only if the source page was read, and it is cited with its URL
- Content pulled from the web enters the prompt as data, never as instructions
The takeawayAn agent that cannot prove what it claims has to say so. That is the rule Velq imposes on the model before it writes a word.