Building an AI sourcer that actually finds the right talent

aless prx — Sun, 10 May 2026 19:57:31 +0000

Spent the last few weekends building an AI sourcing agent. Quick version: feed it a role, it ranks candidates, drafts the outreach. The model layer was the easy part. The data layer is where every project like this quietly dies.

Every "B2B data" provider gives you the same five filters — job title, location, company, seniority, industry — and a giant pool of half-stale profiles. So your "AI sourcer" ends up recommending the same 200 people as everyone else's AI sourcer, with last-quarter's job titles. That's not a sourcing agent. That's a list rental.

I switched the data layer to DataForB2B and the agent suddenly had something to actually reason over. 70+ filters — including the ones that matter for technical sourcing: GitHub repositories, certifications, languages spoken, years of experience, past employers, funding stage of current company. Profiles are sourced live across 60+ public sources, not pulled from a cached export that's been rotting in a warehouse for nine months. Work emails and GitHub profiles come back enriched on the same call. GDPR/CCPA compliant.

Three things I'd flag if you're building anything similar:

Niche filters are the whole game. Anyone can return "Senior Engineer at Series B startup." Returning "Senior Engineer, contributes to Rust async runtimes, was at the founding team of a YC company" is what makes the agent feel smart.
Live > cached. A "match" who left the company eight months ago is worse than no match — your agent confidently emails the wrong person.
Enrich in the same call. Two-step pipelines (find → enrich) double your latency and your failure surface. Get the email and the profile back together.

If your AI agent is only as good as the B2B data API under it, that's where to spend the next afternoon.

One API for LinkedIn, WhatsApp and Email outreach

aless prx — Sun, 10 May 2026 19:54:06 +0000

Building an AI sales agent on the side. The LLM part is the easy 20%. The hard 80% is the outreach plumbing — LinkedIn, WhatsApp, and Email each come with their own auth model, their own rate limits, their own webhook flavor, and their own way of breaking on a Tuesday afternoon.

Three integrations, three sets of edge cases. Before my agent could send a single message, I'd spent a full week on glue code and zero time on the agent itself.

So I stopped maintaining the plumbing and pushed it to LinkupAPI. Same surface for all three channels, managed accounts, a single webhook for replies and checkpoints. The agent just decides who, when, and what — and calls one endpoint.

Three things I'd flag to anyone wiring up an outreach agent:

Multi-account rotation matters more than clever copy. Build it before you tune prompts.
Agents retry. A lot. Make every send idempotent or you'll double-message your best lead.
Treat checkpoints as a normal event, not an error. Pause that one account, keep the run alive.

That's the whole post. If your agent is drowning in outreach integration code, a unified outreach API is the part most people underweight until they've already built — and re-built — the wrong thing.

DEV Community: aless prx

Building an AI sourcer that actually finds the right talent

One API for LinkedIn, WhatsApp and Email outreach