Kjel Halsey Schlemmer Contact
← Back

Quiver

The prospecting tool the hackajob US sales team starts each day in. It finds hiring companies, enriches them, proposes buyers, drafts outreach for human review, and hands approved contacts to the sequencer.

In production FastAPI React Clay Salesforce Salesloft

Results

+26%
Cold pipeline, with four fewer reps
+34%
Meetings booked
38% → 60%
Enterprise share of cold meetings
+69%
Average cold opportunity size

Team-level cold outbound, pre- vs post-launch. Internal Salesforce and Salesloft reporting.

Problem

Account research was spread across a job board, the CRM, LinkedIn, an enrichment tool, and the sequencer. Assembling that context took reps the first hours of each day, and nothing screened new accounts against existing customers or open opportunities.

How it works

1

Find accounts

Scheduled jobs surface companies that are actively hiring, and closed-lost deals hiring again, then screen them against Salesforce customers and open opportunities.

2

Enrich

Company data fills in cheapest-first: free ATS feeds, then Salesforce, then paid providers, each behind a credit cap and a CRM pre-screen.

3

Find contacts

Buyers are proposed per account. Clay verifies emails and phones and returns them by webhook; a resolver files each contact under the right account.

4

Draft outreach

Briefs and draft emails are generated per account with research attached. A quality check passes each draft, retries once, or holds it for review.

5

Hand off

Approved contacts push to Salesloft cadences. Salesforce remains the system of record.

Architecture

\nFind\nhiring signals,\nclosed-lost\n\n\nScreen\nvs customers,\nopen opps\n\n\nEnrich\ncheapest-first,\ncredit caps\n\n\nContacts\nClay verify\nby webhook\n\n\nDraft\nbrief + email,\nquality gate\n\n\nReview\nhuman\napproval\n\n\nSalesloft\ncadence push,\nSF record Every stage is resumable; paid calls sit behind credit caps; nothing sends without human approval.

Key decisions

Salesforce ownership is the source of truth

Accounts, not batches, carry the work history, so the next rep inherits context.

Buyer targeting is learned, not hardcoded

Target titles come from Salesforce meeting outcomes and are refreshed on a schedule.

Model output is validated in code

JSON outputs are parsed and checked; a person approves every email before it sends.

Technical details

Testing
545 tests, with the model mocked
Cost tracking
Every model call logged with model, tokens, and price
Models
Tiered: a small model classifies, a larger one writes
Data
SQLite on a persistent volume, idempotent migrations
Auth
Google OAuth restricted to the company domain, role-based access
Hosting
Railway
Ownership
Designed and built solo; since extended by the team