Leadrin

Leadrin is an AI sales-agent platform for car dealerships. I co-founded it in March 2026 and run engineering as CTO. The product has two halves that share one fact base: a buyer-facing Customer Page that turns a sales lead into one inspectable place to verify the vehicle and contact the salesperson, and a Sales Agent that handles SMS follow-up inside DealerCenter (the dealership's CRM) without taking the salesperson out of the loop. It now runs across 70 dealerships, has processed 60k+ leads, and has driven $50M+ in vehicle sales.

problem

A dealership lead has a half-life measured in minutes. The slow first response, the lost-in-the-feed inquiry, the salesperson buried under thirty other threads: that's where leads die. Existing CRMs are dashboards. They don't reply, they don't book, and they don't ground anything in the actual car. Buyers, meanwhile, get generic "thanks for your interest" templates that don't tell them whether the car is still there, what it costs, or who they'd be working with.

approach

Two products glued together by one fact base.

• Customer Pages are fact-first vehicle research packs. One page per lead. Sourced VIN/inventory facts, real photos with quality filtering, the salesperson's identity and intro video, a direct path to chat or book. The page is the artifact the agent delivers and references.
• Sales Agent lives in DealerCenter SMS. A Playwright browser session opens the dealership's own CRM thread, scrapes replies, drafts grounded responses against the Customer Page's Vehicle Proof, and books appointments only when the buyer confirms an offered slot. It hands off silently to a salesperson the moment the conversation enters negotiation, legal claims, or off-script territory.
• Workflow modes (manual, auto, agent) let dealerships start conservative and graduate into automation at their own pace.
• Agent context engine with per-tenant retrieval and structured tool outputs, so every dealership's agent answers only from that dealership's inventory, pricing, and staff. It has handled 90k+ messages for 850+ users.
• Self-improving evals. Every salesperson edit to a drafted reply is auto-labeled into an eval suite that now holds 8,000+ cases. That suite drives model routing per conversation stage, which cut inference cost by 41% without a measurable drop in reply quality.

trade-offs

• Send through DealerCenter, not Leadrin's own A2P. Outreach lands in the dealership's native CRM thread so logging, history, and compliance stay where the salesperson already lives. Cost: we're operating someone else's UI through Playwright browser agents, which is brittle, and keeping multi-stage CRM workflows under 800ms of agent response latency took real work. Benefit: no parallel thread for the buyer to ignore.
• Grounded conversation only. Every vehicle fact, price claim, or dealership claim must come from sourced page facts. The agent can be conversational, but it cannot invent. When it can't ground something, it hands off. Trade: less surface area, more trust.
• Conservative rollout, then scale. The first live pilot was a single salesperson at a single dealer running a three-touch sequence (initial page, same-day check-in, next-day ask). Every new dealership still starts in manual before graduating. Trade: slower onboarding, but the agent learns where the real handoff edges are before it touches a wider population.
• Humans close in person. The agent gets the buyer informed and booked; the salesperson records personal media when useful, handles handoffs, and shows up to the appointment. We're not trying to remove the salesperson; we're trying to give them their morning back.

status

Live across 70 dealerships after a June 2026 single-dealer launch. The platform has processed 60k+ leads, handled 90k+ messages for 850+ users, and driven $50M+ in sales. Active build: widening workflow-mode coverage, hardening DealerCenter Sync as the dealer count grows, growing the eval suite past 8,000 cases, and extending Customer Page evidence sections.

what i'm learning

The hard part isn't the LLM. It's the integration surface: DealerCenter isn't an API, it's a CRM UI. Most engineering hours go into the scrape/send contract, recovery from session changes, and stage-write idempotency. The agent is only as good as the page it grounds in, and the page is only as good as the inventory data feeding it. Most of the value shows up in the boring layers underneath.

The second lesson is that evals compound. Once salesperson edits started flowing into the case suite automatically, model routing stopped being a guess and cost went down while quality went up. The feedback loop turned out to matter more than any single prompt.