A voice and avatar hiring agent that ran an 800+ candidate funnel
Written by Shahzaib Shoaib, founder and lead engineer
A solo build for a hiring team. An agent runs phone interviews by voice, follows up with AI-avatar videos, then analyzes the answers and hands recruiters a shortlist with clean summaries. Built on n8n, Vapi, ElevenLabs, HeyGen and Claude.

- 800+
- Candidates run through the funnel
- Client
- Hiring team
- Our role
- Solo build
- Industry
- Recruitment · Hiring
- Services
- Voice AgentAI AvatarAutomation
A phone call nobody on the hiring team makes
A candidate's phone rings. The voice on the line is calling about the role they applied for, and it has questions about it. The candidate answers out loud, the way they would with a recruiter, and then the call ends. Nobody from the hiring team was on it.
Later, a recruiter opens a shortlist. Each name on it has a clean summary attached, so the recruiter starts from the people worth a second conversation and never has to sit through every first one.
That is the system we built for a hiring team. It is a voice and avatar agent that runs the top of their hiring funnel. Phone interviews come first, then AI-avatar video follow-ups, then analysis of the answers and a shortlist. The summaries go to humans, and humans decide what happens next.
I built it alone and owned it end to end. The client is private and under NDA, so this write-up keeps the company and every candidate out of it. The images on this page are mock-ups we made for it.

Five tools, one candidate at a time
The stack is n8n, Vapi, ElevenLabs, HeyGen and Claude. Each one has a narrow job.
Vapi runs the phone calls, and the agent speaks with an ElevenLabs voice. HeyGen produces the avatar videos for the follow-up round. Claude reads the responses and does the analysis the shortlist is built on. n8n sits under all of them and moves each candidate on to the next step when the last one is done.
The orchestration is the part people undersell when they describe a system like this. The stages don't happen in one sitting. A call ends, its answers need analyzing, a follow-up has to go out, and the candidate's record has to say where they are at every point in between. If that record is wrong, a person falls out of the funnel and nobody notices.
So the routing lives in n8n as ordinary workflow logic. Someone on the client side can open it, see which step a candidate is on, and change a rule without touching a prompt. I wanted the models doing the parts that need language, and plain logic doing the parts that need to be predictable.
Why the phone call comes before the video
The order of the rounds is the choice I'd defend first if someone pushed on this design.
A phone call asks very little of a candidate. They pick up and talk, which every applicant already knows how to do. A video round asks for more of their attention, so it belongs later, once the candidate has a reason to give it.
The second reason is comparison. The phone interview works from set screening questions for the role, so every candidate on a call is answering the same things. That is what lets the analysis put two people side by side and have it mean something. If the first round wandered, the shortlist would be ranking different conversations against each other.
The follow-up is where the system can afford to vary. By the time someone reaches the video round, their phone answers are already on record, and that is the natural place to go deeper. That is how it was designed. I don't have data on how much the follow-ups differed between candidates, so I won't claim a number for it.
The trade-off is real. A candidate's first conversation with this company is with a machine, and some people won't want that. Two automated rounds also give a candidate two places to drop out instead of one. I don't have a count of how many people left at either point, so I can't tell you what that cost this client in good applicants.
The number I can stand behind is the size of the funnel. I should have counted every stage from day one.
What the recruiter actually reads
The output of all this is small on purpose. A recruiter gets a shortlist, and each person on it comes with a clean summary.
I went with summaries because nobody on a hiring team has time to listen to every first-round call, and a raw transcript is just a recording you have to read. What a recruiter can use is a short page on what the candidate said about the things the role needs, written so they can scan several of them before deciding who to call back.
That summary is a starting point for the recruiter's judgment. The design keeps a person between the shortlist and any offer, and I think that is the right place for the line. The agent's job is the first-round work, done the same way for each candidate. Deciding who to hire is a different job.

What the 800+ counts
The figure I can put my name to for this project is an 800+ candidate funnel.
It describes the size of the funnel the agent ran. It isn't a count of completed interviews, and I'm careful about that because the two are easy to blur. A funnel narrows by design. Its size tells you how many people went in at the top and says nothing about how many finished each stage. I don't have a verified count of completed calls or video follow-ups that I'm willing to publish, so there isn't one here.
There are other gaps too. I have no measured figure for recruiter hours saved. I have no data comparing the agent's shortlists with what a recruiter would have picked from the same pool, so I can't tell you how often they agree. Those are the numbers a buyer should want, and I'd rather print a smaller one that holds up.

What I would build in from day one
I should have built the reporting into the workflow at the start. Every stage in n8n already knows when a candidate enters it and when they leave. A plain count per stage, agreed with the client while the project was live, would let me tell you today how many people finished a call and how many reached the video round. It is a small piece of work at the start, and its absence is the biggest hole in this write-up.
The second thing is a check on the analysis. Before trusting a shortlist at volume, I'd take a sample of candidates, have a recruiter rank them blind, and compare those picks with the agent's. Until that is done, the summaries save reading time and the ranking behind them is unproven.
The figure on record stays where it was when I wrote this: an 800+ candidate funnel, with the completion counts still to be confirmed.
- Vapi
- ElevenLabs
- HeyGen
- Claude
- n8n
A screening agent earns its place by asking every candidate the same first-round questions and handing a person a summary they can act on. Put the step that asks least of the candidate first, keep the routing in plain workflow logic, and count every stage from the first day, because later you'll want to know how many people finished each step and the top-of-funnel figure won't tell you.

