'/%3E%3Cdefs%3E%3ClinearGradient id='paint0_linear' x1='0' y1='0' x2='0' y2='2160' gradientUnits='userSpaceOnUse'%3E%3Cstop/%3E%3Cstop offset='.1' stop-color='%23000' stop-opacity='0'/%3E%3Cstop offset='1' stop-opacity='0'/%3E%3C/linearGradient%3E%3C/defs%3E%3C/svg%3E%0A)
AI-driven ad approvals
for 
A challenging situation
In 2012, Yahoo Japan had a problem. To comply with strict government regulations, each ad request had to go through a manual verification process before it could go live in production.
The process involved a rigid hierarchy of managers and employees to perform the review and approval tasks for each request, causing delays and numerous lost revenue opportunities for Yahoo Japan.
The process involved a rigid hierarchy of managers and employees to perform the review and approval tasks for each request, causing delays and numerous lost revenue opportunities for Yahoo Japan.
The big consulting firms
Yahoo Japan brought in several global consulting firms to come up with possible solutions. All firms gave them the same answer—hire a larger workforce, with some estimates close to 10,000 contractors to deal with the scale and scope of the problem.
The SiliconMint approach
We looked at the problem in a different light, for two reasons:- We are obsessively customer focused.
- Our technology expertise is both deep and wide. This allows us to rapidly perform R&D on what's technically feasible.
We got to the heart of what Yahoo Japan executives wanted:- Avoid increasing headcount. Hiring thousands of contractors would destroy the division's tightly knit corporate culture built carefully over a long period.
- The division needed to provide an SLA on ad request approvals, such that other departments could better rely on them.
With this in mind, we proposed building a system that:- Uses artificial intelligence (machine learning).
- Allows Yahoo Japan to keep their workforce the same size.
- Allows ads to go live in near real-time.
Technical architecture
Artificial Intelligence Core
Determines the best worker for each task in real-time, while guaranteeing accuracy & SLA.
Visual Task Editor
Drag and drop, visual construction of verification tasks.
Workflow Engine
Transactional engine to hand out and correlate async results.
Real-Time Workload Distribution Cluster
with custom cache
with custom cache
Custom software to support key business objectives
The key business objectives were clear—keep the workforce the same size to maintain company culture and allow for ads to be approved in near real-time while adhering to government regulations.
To meet these goals, we analyzed off-the-self technologies and solutions and came to the realization certain systems would need to be custom built. With Yahoo Japan's approval, we built the following custom tiers.
To meet these goals, we analyzed off-the-self technologies and solutions and came to the realization certain systems would need to be custom built. With Yahoo Japan's approval, we built the following custom tiers.
Real-Time Workload Distribution Cluster
To meet the required latency & throughput SLA, data was kept in memory on a custom-built cluster on top of Hazelcast.
Artificial Intelligence Core
A mechanism using binomial distribution along with worker history to train and execute decisions for which worker should receive a given task in real time.
Visual Task Editor
A web-based application to visually define the user interface for any classification task, along with visual definition for how and where data would connect in the Workflow Engine.
Workflow Engine
An in-memory engine to maintain worker state, taking into account input from the Artificial Intelligence Core to drive subsequent steps for each task until the required accuracy & SLA were delivered as output.
The timeline
With the architecture and development plans in place, we set an aggressive schedule to deliver the first version of the system to Yahoo Japan in nine months.
We delivered on time and launched the first version at Yahoo Japan in just nine months. We then spent the next five months improving the system and getting it ready for production:
1. Tuning the artificial intelligence core (various parameters)
2. Gathering feedback from users to improve the Visual Task Editor
We delivered on time and launched the first version at Yahoo Japan in just nine months. We then spent the next five months improving the system and getting it ready for production:
1. Tuning the artificial intelligence core (various parameters)
2. Gathering feedback from users to improve the Visual Task Editor
The results
The project was a runaway success—Yahoo Japan was able to increase the speed with which ads are approved by over 1,600% (16x). But, perhaps more importantly, they were able to achieve their business objectives without increasing the size of their workforce or destroying their company culture.
The executives in charge of bringing us in as an external team were promoted and internally praised many times over for this fact alone.
The executives in charge of bringing us in as an external team were promoted and internally praised many times over for this fact alone.
Before
After
Benefits
AI Engine
–
0 - 60ms / task
Immediate task
assignment
assignment
Human Work
1 - 3days / step
30 - 60sec / task
More than
144x faster per task
144x faster per task
Workforce Size
~200people
~200people
Workforce size
remains the same
remains the same
Entire Workflow
24 - 72hours
2 - 3hours
1,600% (16x) faster
to approve an ad request
to approve an ad request
Get in touch
To discuss your development needs
Attach files
Send me an NDA
Attach files
or drag & drop files here
Send me a SiliconMint NDA
Send