Smart Web Agency, for a recruitment client under NDA

27,431 applications screened by a pipeline I built in 18 hours

A recruitment client of Smart Web Agency wanted AI screening added to the hiring pipeline it already had. I built it in 18 hours of total time, on Laravel, Redis queues, LangChain and Gemma 27B. It worked through every application in nine days, the client reported almost 72% less manual work, and its hiring round moved from July to April.

To build the whole pipeline: 18h
Applications screened: 27,431
Less manual work, reported by the client: ~72%
Annual hiring round: July to April

How the pipeline works

Screening pipeline

One hiring round, every application

27,431 applications

Built in 18 hours of total time

Laravel

MySQL

Redis

LangChain DeepInfra Gemma 27B

One application, end to end

Laravelreads the application Redis queueholds the job for a worker LangChainbuilds the prompt, parses the answer Gemma 27Bon DeepInfra, screens it MySQLstores the result for recruiters

One batch run, nine days

0 screened

Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 Day 8 Day 9

The hiring round, on the calendar

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

July to April, about two and a half months earlier

The brief

A hiring round of 27,431 applications

The client runs one big hiring round a year. It wanted an AI screening layer added to the hiring pipeline it already had, so every application gets screened without changing how its recruiters work.

The build

Eighteen hours, start to finish

I built the whole pipeline in 18 hours of total time: Laravel and MySQL for the application data, Redis queues for the work, LangChain around the model, and Gemma 27B served on DeepInfra.

How it works

Every application is a queued job

Laravel puts each application on a Redis queue. A worker picks it up, LangChain builds the prompt for Gemma 27B on DeepInfra and parses the answer, and the result goes back into MySQL, next to the application, for the recruiters to use.

The run

27,431 applications in nine days

The pipeline screened every application in one batch run of 9 days. Because each one is its own queued job, the run keeps going on its own and picks up where it left off.

The result

Less manual work, and an earlier hiring round

The client reported almost 72% less manual work. With screening no longer the slow part, its annual hiring round moved from July to April, about two and a half months earlier.

Before and after

Manual screening work, as reported by the client

Before

About 72% less

Before After

Annual hiring round

Before

After

Before, the hiring round ran in July. After, it runs in April.

The work

Built at: Smart Web Agency, for a recruitment client under NDA
Build time: 18 hours in total
My part: The whole pipeline: queueing, the LangChain layer, the model integration and storing the results

Laravel
MySQL
Redis
Queues
LangChain
DeepInfra
Gemma 27B

The story behind it: they asked for AI resume screening in a day.