Advertisers via plain Google
Instead of fighting the Google Ads API and its limits, it pulls advertisers the way a user would — searching Google for the client’s keywords and collecting whoever shows up.
← All cases Source Gather Extract Qualify
AI · Sales ops
AI that turns keywords into call-ready leads
An AI pipeline that turns a list of keywords into qualified, call-ready leads in the CRM — it finds the companies, gathers their contacts, structures the mess and scores every lead, with nobody touching it in the middle.
The problem
The client’s sales team burned hours on the same loop: search Google for a list of keywords, see which companies advertise there, hunt down their contacts, open their sites for more, then judge whether the lead is even a fit. About half the effort was wasted — the lead turned out off-target only after all the prep.
They needed a machine that does the whole grind and hands sales only finished leads — contacts, a description, and a “worth a call” verdict.
What I built
A six-stage pipeline from keyword to CRM record. The stages are independent, so a mid-chain failure only re-runs that one step.
Contacts & site enrichment
For each company it collects phones, emails and office addresses from the results, then visits the site to pull a business description, key people and product pages.
spaCy structuring + normalization
Snippets and HTML arrive in wild formats. spaCy cheaply pulls structured entities — people, official company names, short summaries — and a separate layer normalizes them to one shape before anything moves on.
LLM scoring → Zoho
Each lead goes to an LLM with the client’s criteria: a fit or not, hot or cold, what the hooks are. The verdict plus contacts and description is written straight into Zoho CRM — the manager just calls.
What made it hard
- Heterogeneous sources — Google snippets, site HTML, meta tags — all in different shapes. Normalization grew into its own rule-heavy module.
- Balancing spaCy ↔ LLM: all-LLM is slow and pricey, all-spaCy misses meaning. The fix was a cascade — spaCy extracts entities first, the LLM judges only what needs understanding.
- Per-client prompt engineering: each client’s qualification criteria differ, so the LLM prompt is assembled from a couple of settings — swap the client, re-tune the qualifier, no code changes.
The result
- Sales receives finished, qualified leads directly in Zoho CRM.
- The full chain — keyword to CRM record — runs with no manual step in the middle.
- Managers stop hand-vetting every company — they just call.
What I took awayAll cases →The spaCy + LLM cascade is a strong pairing: fast, deterministic tools do what they’re good at (structured extraction), and the LLM is reserved for what genuinely needs context. I now reach for that split on most text-data projects.