Stryke
An AI sales platform for car dealerships. Agents work inbound leads over SMS, negotiate trade-ins inside each dealer's limits and book the appointment, with every hard rule enforced in code.
Ryne Benson
Product · Automotive · 2 min read
Dealers get leads from everywhere, and each one needs a fast reply, a qualifying conversation, a fair trade-in number that stays inside the store's limits and a booked appointment. All of it has to respect texting consent.
Stryke hands that work to agents. Axel talks to sellers and negotiates, Nova talks to buyers and searches the lot, and Insights answers dealer staff about sales history, inventory and the calendar. Staff work the pipeline in a dealer app, and a separate admin app and public trust center sit beside it.
How it works
A lead comes in
Leads arrive as ADF emails through SES and S3 or through an intake webhook, and land in the dealer's pipeline.
Consent before the first text
Email-first sources get a single email with a consent link. Texting opens only after the customer opts in, and a snapshot of that page is kept as write-once evidence.
Every message passes a send gate
Before anything goes out, a gate checks sending hours, suppression and opt-out lists, kill switches, a send cap and the message content itself.
The right agent takes the thread
Sellers talk to Axel, who gets an offer plan from the pricing engine, walks an offer ladder and books an appraisal. Buyers talk to Nova, who searches inventory and books a test drive.
Decisions worth explaining
The model never computes a price
The maximum offer comes from a pricing engine that only looks at the vehicle and the market. Tool inputs are allowlisted so consumer attributes are rejected, every computation is saved so it can be reproduced, and tests fail if the TypeScript and Python engines drift apart.
Guards on every SMS reply
Replies are buffered and the customer-facing text is pulled out of tags. Emoji, odd punctuation or leaked narration trigger a regeneration, and a validator blocks offers over the limit, claims to be human and pressure tactics by handing the thread to a person.
Evals against the deployed agents
Scripted scenarios run against the real agents and are graded twice: deterministic checks replay the offer plan to verify every number, and an LLM judge that fails closed covers the rest.
Observation records
Each negotiation keeps a permanent record of decisions, counterfactuals and outcomes, with randomized control groups per dealer, so changes to Axel can be measured instead of guessed.
Personal data stays out
Every Lambda redacts phone numbers, emails and SSNs from logs, CI fails on whole-payload logging, and uploads with personal data in any cell are rejected.
By the numbers
- 3 agents
- 4 mcp servers
- 9 skills
- 8 eval scenarios
The stack
- Interface. Next.js, shadcn/ui, AWS IoT, Web push
- Agents. Claude Agent SDK, Bedrock AgentCore, Model routing, Evals
- Tools. FastMCP, AgentCore Gateway, Cognito M2M
- Data. DynamoDB, DynamoDB Streams, S3 Object Lock, SSM
- Platform. SST, Lambda, SQS, SES, WAF, Cognito, ECS
- Delivery. GitHub Actions, semantic-release, Personal stages