Built a four-agent AI SDR platform that researches, qualifies, emails and books meetings with a human in the loop

The problem

Sales development is a grind of research, list-building, first-line writing, follow-up and reply triage. Small teams either hire SDRs they cannot afford or let outbound stall. Off-the-shelf AI email tools handle sending but leave the hard parts — deciding who fits, keeping a human in control of what the AI says, respecting suppression lists and keeping an audit trail — to the user.

We wanted a product that ran the whole loop, from a raw CSV or an Apollo search through to a proposed calendar slot, without removing human oversight. That meant explainable qualification rather than a black-box score, approval on the first AI-sent reply, a real pre-send compliance gate, per-workspace AI budgets, and enough operational rigour (role-based access, one-account-one-workspace, tamper-evident audit logs) that a paying customer could trust it with their sending domain.

What it cost

Without a product like this, every outbound campaign at Techvera and its clients depended on manual research and hand-written sequences, and there was no repeatable way to package that work for other businesses. Building on a generic email tool would have meant no control over qualification logic, no conversation handling, and no compliance guarantees.

What we built

VeraiSDR: a controlled four-agent pipeline. The Research Agent enriches sparse leads using Tavily and Serper; the Qualification Agent scores each lead against a workspace ICP and returns explainable pass, review or reject decisions; the Outreach Agent drafts sequenced emails; the Conversation Agent classifies replies and routes positive ones into an inbox, proposing calendar slots. Leads arrive by CSV upload with header mapping and dedup, or via bring-your-own-key Apollo search with keys encrypted in Vault.

Around the agents sits a full SaaS shell: workspace signup with seat caps, saved lead views, lists and segments, campaign step editor, per-agent model tiers and daily budgets, server-side RBAC on every controller, hash-chained audit logs, Stripe billing, and a human-in-the-loop approval gate on the first AI reply. Thirty-one MVP capabilities, 75 frontend routes and 39 backend modules, with 314 verified migrations and a Playwright suite covering the core flows.

What changed

The MVP is go-live ready in dry-run mode: builds are clean, unit and migration checks pass, and the send funnel is idempotent with a real pre-send compliance gate. Live external sending, billing and CRM sync are pending production credentials, so outcome figures are not yet available.