Pole Position Agency.
My roleOwner · built the platform and the AI
Pole Position Agency is my own marketing agency for car washes, set up in 2023 as the first project after Exito and Place Here merged. In thirteen months I built its platform: one database where till, number plate cameras, loyalty and ads come together per location, with an email agent on top that answers customers from a knowledge base. Every reply is checked against its sources. When money is involved or the agent is unsure, a person decides.
- What I did
- As owner I designed the platform and built it with AI: the data per site, the dashboard and the email agent.
- 1M+
- customer emails sent automatically
- 96,417
- vouchers redeemed
- 132,000
- customer profiles
- 39
- locations across 17 businesses
The monthly view of what the agents did on their own: emails answered, new questions added to the knowledge base, links found between weather and visits.
- Number plate recognition
- Savings programme
- Come-back voucher
- Chatbot
- Facebook and Instagram
- Social media
- Weather
- Customer value
- Revenue
- Tasks and reports
One database for every car wash, the AI on top of it, and a person on every decision that costs money.
- Mail comes in through a webhook
- Filters, rate limit and idempotency
- Triage, member lookup and thread history in parallel
- Hybrid search: pgvector and Dutch full-text, fused with RRF
- A small or a large model writes the reply
- Every claim checked against the sources
- Send, ask the customer, or hand it to the owner
Hybrid retrieval
Vector search alone misses prices and number plates. Cosine similarity and Dutch keyword rank are fused with Reciprocal Rank Fusion, so an exact keyword hit survives even below the similarity threshold.
Checked against its sources
Every concrete claim in a reply is checked against the retrieved answers and rewritten once if needed. Still ungrounded means a draft for the owner, and a question without a source never gets a reply on its own.
Knowledge that grows
When the owner answers, a curator model turns that answer into a reusable entry without personal data. It waits for approval, and a gate in the database itself rejects IBANs, plates, phone numbers and email addresses.
Cost-aware routing
A small model triages and answers simple mail; complaints, attachments, several topics or another language go to a larger one. Measured: $0.0068 of AI per email thread on average.
Patterns in the visits
Number plate scans per location are correlated with each kind of weather, a drop of a quarter below the 30-day average raises an alert, and customers are scored on recency, frequency and value.
Isolated and safe to show
Every query is bound to company and location by row-level security. The public demo runs on a token that Postgres forces into read-only transactions.
Who is coming in?
Cameras read the plate on arrival and match it to the customer profile: new or returning, how often, with or without a purchase. So the right offer is ready.

The right model per email
A small model triages and handles simple emails; complaints, attachments and harder questions go to a larger one. On average an email thread costs $0.0068 in AI.

The car wash in your pocket
Owners check visits, members and campaigns on their phone, and work through the week's tasks between two cars.




The merger's first project
Set up as Common Grounds' first project. The platform stood within thirteen months; today it runs for 39 locations across 17 businesses.
Per location or per group
Each location has its own view, groups of locations sit side by side, and every number can be brought down to one site.