Experiment 1 · trdi/trady
What it was, and what I learned.
The starting hypothesis was that public data could do double duty: power a product and serve as its go-to-market channel. In the right market, publicly observable signals could be turned into actionable intelligence that helps businesses identify prospective customers and gives the product a built-in path to distribution.
I explored this across traditional trade and field-service verticals, including HVAC, cleaning, fire safety, pest control etc. This became trdi: an attempt to build an autopilot for local lead discovery that turns fragmented data into prioritised prospects and call-ready sales briefs, rather than another static lead database.
Pest control was the first vertical I took forward. NYC publishes restaurant health inspection records, including violations that can indicate potential pest issues. Combined with reviews and other data sources, these records offered a way to identify and qualify prospective accounts for pest control operators (PCOs). The proposition was to give operators a better way to identify restaurants worth pursuing, with account-level intelligence to support their outbound sales.
I built the product using Claude Code, Supabase and Vercel. The pipeline ingested NYC restaurant inspection records, then enriched the data with Google places API and reviews, Yelp ratings and other public and paid API sources. It generated restaurant-level risk profiles, with additional analysis across building & neighbourhood clusters and 311 complaints. Daily agents handled ingestion, enrichment, risk-score recalculation, profile refreshes and tailored prospecting briefs for individual PCOs.
The idea was to move beyond static lead lists. Instead of asking an operator to research every account manually, the system would continuously surface prospects and explain why they might be worth pursuing.
I built two landing page variants but never drove traffic to them. Rather than spend time optimising a funnel before establishing demand, I took the proposition directly to prospective buyers through sequenced cold email. Some messages were short pitches; others included a one-page brief showing the signals we'd found at specific restaurants. The hypothesis was that PCOs would recognise the opportunity, contact the restaurant and convert it into new business.
The result was one phone conversation, a couple of email threads and no paid conversions.
That's a small sample, not enough to conclude that PCOs wouldn't pay for this. But it was enough to expose weaknesses in the assumptions behind the product.
The first was the relationship between the signal and the buying decision. An inspection violation could indicate a pest problem, but it didn't tell me whether the restaurant was looking for a new provider. Many restaurants already had a PCO. There was no reliable way to infer dissatisfaction with the incumbent, switching intent or an active procurement process from the data I was using.
I'd built a scoring model to identify accounts with a potential problem and treated that as a proxy for purchase intent. Those are two very different things.
The second issue was customer economics. I hadn't established how much an incremental restaurant account was worth to a PCO, what conversion rate would justify the acquisition effort, or what operators would pay for the intelligence. Restaurants could also be demanding accounts to service, with callbacks, scheduling constraints and price sensitivity. Even a relevant lead wasn't necessarily a valuable one.
Then there was distribution. Public records made restaurants identifiable, but reaching the owner was another problem. Gatekeepers and incomplete ownership information made direct access difficult. My parallel outreach to restaurants ran into the same constraint. The product could identify an account, but the path from that insight to a signed contract remained indirect: the PCO still had to reach the decision-maker, establish a need, compete with an incumbent and close the business.
The technical system worked as intended. It ingested, enriched and scored accounts, refreshed profiles and generated tailored prospecting briefs. But I was spending inference and development costs improving the intelligence without evidence that it would change the buyer's behaviour. More enrichment could produce a better account profile. It couldn't establish whether that account was in market or worth acquiring.
In hindsight, I'd moved too quickly from an observable signal to a product proposition without validating the commercial assumptions in between. The experiment showed that public data could support both product functionality and prospect discovery. It didn't establish that the resulting intelligence was differentiated enough to command a price, or that the proposed distribution channel could convert it into revenue.
The broader thesis still interests me. I'm working on an agent to help identify and evaluate niche opportunities using signals from multiple sources. The goal is to get beyond interesting datasets and technically feasible ideas, and systematically assess where existing solutions fall short, whether buyers have a material problem and whether there's a credible path to monetisation.
I'm also exploring a services-first approach. Rather than build SaaS upfront and hope demand emerges, I want to solve a specific problem for 5–10 paying clients using AI, learn the workflow through delivery and identify what can be productised into software. That should provide a much stronger feedback loop around willingness to pay, operational complexity and the parts of a workflow that are genuinely repeatable.
Before committing to the next build, I want evidence of a real buying trigger, a viable route to the decision-maker and customer economics that justify solving the problem. Those are the assumptions to validate before investing in automation, not after.
The opportunity isn't necessarily where the data is richest or the automation is most impressive. It's where the signal translates into a business outcome someone will pay for.
On to the next one.
