// case study · 2026
K9 Signal Radar
An AI prospecting engine that finds working dog programs in the news and US procurement, finds the right person, and starts the outreach.
- Role
- AI Engineer
- Company
- DogBase
- Type
- AI & automation
- Outcome
- 25 agencies checked for a contact
01
The problem
DogBase sells to police K9 units, search and rescue teams and detection programs. These buyers don't fill out lead forms, and they rarely search for software until a budget is already approved. The early signs are out there, just scattered. A local paper covers a new K9 unit, a business donates a dog to the sheriff, a county posts a procurement notice. I was finding these by hand, a few hours a week, and most of what I found was noise.
02
Constraints
The sources share nothing. News comes as RSS, and SAM.gov has its own API and format. The outreach side has its own rules. I moved from Apollo to Instantly, where a verified email costs credits, the plan holds 1,000 contacts, and a bad batch of addresses can hurt every mailbox I send from. Police and sheriff's offices are also hard to find emails for: so far about 3 in 7 real lookups have found one. It all had to run on its own every day, cost close to nothing, and never buy junk or contact the same agency twice.
03
Decisions & trade-offs
A fixed pipeline with the AI doing two small jobs, instead of an autonomous agent.
- Plain code does the fetching, filtering, deduplication, and outreach writes. Claude does two jobs: Sonnet reads a signal and returns structured JSON saying what kind of signal it is, which organization it's about, and whether a person is named. Haiku searches the web for the K9 contact at each new agency.
- A keyword check runs before the model. The working dog world has its own vocabulary, so most noise is gone before you spend a single token.
- Credits are spent on purpose. Searching for people is free, so the workflow buys at most one contact per new agency and five per run, and it stops before the contact limit.
- A person has to match before anything is bought: a K9 or command title, the right organization and, for US agencies, the right state. That last rule exists because the first live run picked the sheriff of Jefferson County, Pennsylvania for a story about Jefferson County, Missouri.
- Contacts are a budget too. Bounced and unsubscribed leads are removed every day, and leads who finish with no reply move to a shorter second sequence after 60 days.
- Every signal and every organization is remembered in n8n Data Tables. That stops repeats and lets a failed run retry the next day.
- It collects every day and reports twice. A daily email shows what happened, and on Monday an agency that appeared in three stories shows up as one card with three sources.
- Why not an agent? Because the steps are known and the same every day. An agent would rediscover them on each run, read everything instead of a filtered slice, and could spend credits or email people in ways I can't predict. When something looks wrong here, I open the run and see exactly which step did what.
04
What I built
The whole workflow in n8n. Two live sources: Google News in its US and UK editions and SAM.gov canine notices. EU TED tenders and the EU Funding and Tenders portal are built but switched off. On top of those, a shared format for every item, the keyword gate, the Claude prompt and the parser that reads its answer. Grant calls are routed to me, since they're for DogBase rather than sales leads. Everything else goes through Instantly: Claude finds the K9 contact on the web, then a free preview of who works at the agency, one verified email bought, then a four-email sequence whose wording and links change for police, search and rescue, detection and service dog teams. Separate daily jobs clean up bounces and unsubscribes and move finished leads to the second sequence. Then a daily report and the weekly brief, with every item explaining why it did or didn't become a contact.
05
Where it is now
It runs every day on US and UK signals. I moved it from Apollo to Instantly in November 2026, and its first live run checked 10 new agencies and added 1 contact. So far it has checked about 25 agencies and put 5 people into the new sequence. I later moved the AI from DeepSeek to Claude, because web search named a K9-specific person at 6 of 10 test agencies where Instantly managed 1. Finding the signal is no longer the hard part. Finding a working email for the right person is. The same structure powers my own fractional work digest, which I wrote about in Why my AI lead generation is a pipeline and not an agent.
// stack
- n8n
- Claude API
- Instantly API
- n8n Data Tables
- Gmail
- Google News RSS
- SAM.gov API
Need this kind of engineering?
Book a call, or send a note about what you’re building.