Cockpit.
I work a sales patch by day. So I built the intelligence layer the job should have come with.
The day job is business development for — a phone, a patch, a number to hit. The standard tooling tells you who to call. It does not tell you what is true about them, what is changing in their world, or what you would say in the first thirty seconds that earns the next five minutes. So I built that part myself, on my own stack, on my own time.
The core is a two-stage reasoning engine. Stage one gathers evidence about an account — filings, signals, hiring, tech estate, market movement — and refuses to hold anything it cannot source. Stage two synthesises strategy from that evidence: an eighteen-section dossier reasoning from business to strategy to operations to problems to opportunity, with the top play scored across nine dimensions. Every claim carries a link and a confidence tier — verified, strong, hypothesis, speculative. If the system is guessing, it says so in writing.
Around the engine sits the working surface: an “arm me for this person” generator that researches a target and drafts the full kit — call script, messages, follow-ups, battlecard, a personality read with its own falsifiers. A playbook matrix crossing eighteen personas with fifteen business problems, tagged by industry. Model routing keeps the economics sane: the small fast model classifies and retrieves, the mid model drafts, the deep model does synthesis — every call logged with its token cost.
Lately it grew two faster ways in and one that needs no hands at all. Quick Assist takes a single lead — pasted in, or read straight off the CRM record or professional profile you already have open by a browser extension I built to sit beside them — and returns the best-fit play, its positioning, and an opener in one step, no full dossier required. The extension is deliberately narrow by design: it acts only when you click it, reads only the company and title on the active record, holds no standing permission to roam, stores nothing, and never sends a person’s name anywhere. It’s installed and tuned to my own patch now, and it’s the part I reach for most — the whole engine, one click from the record already on screen. And a scheduled weekly run briefs itself — producing a fresh dossier and a ranked target list on its own each week, waiting for you when you sit down, still with nothing sent.
Then it stopped being only mine. I demoed it to the team, worked through the feedback, and shipped a shared version — role-gated, with each person’s work attributed under their own login — alongside a guided walkthrough so nobody needs me in the room. And the most interesting move is the most recent: the core engine is now exposed as a service — a gated, rate-limited endpoint other parts of the business can call from their own tooling, each caller on their own key, the reasoning method staying server-side. Built, deployed, and deliberately dark until it gets the internal sign-off it deserves. The same engine is already sketched as the brain behind a website concierge for inbound enquiries. What started as one person’s edge is turning into infrastructure — my intelligence, made callable.
Two design rules hold the whole thing. It never sends anything — every output is prepared for a human to read, judge, and deliver; the system has no send button by design. And it never touches what it should not — no scrapers, no session automation, hard ceilings on every data source, every metered lookup logged. Intelligence, not automation. The judgment stays human.