SATVIK JAGANNATH:
Vitra.ai is
a tool which you
can use to translate
videos, images,
podcast, and text to
50 plus languages with
just one click using
artificial intelligence.
Usually, the process
of translation
in a manual scale is
cumbersome and slow.
Using Vitra, the speed up
is close to 100 times.
The whole process is
over 80 percent cost
effective than any manual
process in the world today.
One of the things with
our own custom models that
we had built for all the
speech-to-text engines and
pipeline was first of all,
the data timing,
there is a lot of
effort that goes into
training the model,
where you get the data
in place and stuff.
That itself is very
cost intensive,
and it's very time consuming
used to take three months,
six months cycles for us to
get a model into and train.
But when we just
did a small POC,
like I said, with Azure
Cognitive Services,
the accuracy of the
punctuations, for example,
was much better,
30 percent better than what
Vitra's own custom models was.
The whole ripple
effect of AI models,
which yielded us overall
of more than 30,
40 percent, better
accuracies and
translation systems and
stuff like that, eventually.
Overall, this yielded
impressive results,
while we explored
in special system.
The cost was reduced by 2.5x,
which was a huge
infrastructural overhead
was broadened down for us,
and it was just a seamless,
simple API integration,
which hardly took like
30 minutes for us
to integrate with
Azure services.
That was the biggest benefit.
We didn't have to break
our heads as to how
to scale our infra Cloud,
the servers, basically,
GPU and CPU servers.
Why, because it's an API.
During the initial days,
some of these speech protects,
especially that system, we
had a certain customer
segments that we
were working with.
They needed custom training
enough for their own data set.
That's when we jumped on to
building our own AI models.
But then we realized eventually
that we couldn't manage
these models. It was huge.
As I said, the
operational complexity,
the infrastructure complexity
was just scaling up.
As the company scaled, when
you start off as a start up,
you want to do everything.
But only then you realize
that uploading a lot of
the things that you
don't want to do at
scale would actually
yield better results.
That's what we felt.
That's when we evaluated,
like a dozen of special
tech systems and
Azure speech protects felt very
effective and efficient
when it comes
to the output that it gave,
the output that it generated and
the speed and the overall cost.
It was not just one factor
that helped us make
this decision.
This decision was
reviewed by hundreds
of people inside the
company and contractors.
Hundreds of people
reviewed and it was
a blind date like black box.
You could just see
the datasets and
you could choose which was
the best performing one.
That's how we derived in Azure,
which, I think, almost got
over 80 percent of the words.
In our analysis that
we made over 30 days.
That was impressive.
When it comes to managing Vitra,
the biggest problem was
how to cater to scale,
the peaks, basically.
Suddenly, you had a
customer who wanted to
use hundreds of hours
over a period of let's
say three hours time.
They wanted the
output, and you have
to break your head and
figure out how to scale
for that period of time.
Though there is auto
scalability, all of that,
but the cost is
still much higher.
Unlike a simpler Vitra,
which is CPU driven,
and you have a load
balancer in front,
when you have a
combination of infra.
We had a lot of stuff on CPU,
a whole bunch of stuff on GPU
and all the data being
managed across the place.
This became component again,
it was multi Cloud
in the first place.
Where would you place
the load balancer?
That was the problem.
That's when we figured out,
managing it this
way is impossible.
Obviously, we initially
brought everything to Azure.
Again with the load
balancer approach.
The whole point is Azure
based load balancer wouldn't
help because always the GPS and
the CPS will be hitting
100 percent utilization.
Because it's always full.
There's no point in load
balancing such systems.
That's when we realize in
a let figure out how
the world works today.
That eventually led
to these decisions.
But yes cost was a factor,
no doubt when we were evaluating
because this helped us improve
our own grass margins
significantly in
the product which we
eventually could even
pass on to the customers
in a certain way.
With a set of customer group.
We did A/B testing
without informing
them what we are doing
internally, obviously.
Then we passed on
my Azure Cognitive
services model
to a certain set of enterprises,
and the others used our
custom models the way it was.
They figured out that
the set of customers
who were using Azure
Cognitive services,
the speech-to-text engine had
three to four much
better improvements.
The amount of time it took
for the process to complete,
let's say, a video
translation process to
complete was much better.
I think it was 50
percent faster than the
previous Vitra's
speech-to-text engine.
The second part was
the overall accuracy
when it comes to the
translation, was significant.
As I said earlier, the
punctuation was much better,
which gave us 30 percent
better contextual accuracies
when it comes to translation.
Overall, it was both
the punctuations and
the translations that led
to impressive results.
At the end of 15 days,
30 days, for trial,
we went back and asked like,
we've observed that your
overall experience has
improved the way you use
the product for the
videos and stuff.
They agreed? They wanted to
know if there was an
update to the product.
We passed out survey,
try to figure out
what was better.
We helped them figure out
that this was the same thing
that you did earlier and
this is how it was improved.
Overall customer
satisfaction score
itself improved by
over 23 percent over
a period of three months
of analysis that we
did after switching
to Azure speech-to-text
which is huge for us.
That's how things transformed.
Now I think we overall have
a much better product.
Obviously, thanks to
Azure and a whole
infrastructure plus
the cognitive services
and the applied AI
infra and this is very helpful.
We're also planning to
use the custom training
of Azure Cognitive Services
where we're able to put
out a lot of our own training
data and stuff and train on
top the existing
cognitive models as well.
Explore the services that exist
today because if you're
starting from scratch,
if you're starting from zero,
and if someone is already
there obviously they're
in a much better place.
Obviously using pre
existing services.
Even if you want to
train on custom data,
and all the amazing stuff that
start ups you want
to do all the time.
But fixate yourself on something
that's an existing solution
because that would be robust.
Your go to market,
would be very, very fast.
It took a year-and-a-half
for us to build
this whole infra.
Using, as I said, it took
half an hour to integrate
speech-to-text of a jar.
As I said, after all
of this process of
year and a half worth of
analysis, experimentation, etc.
We felt Azure is
the best solution.
What I would suggest is,
whatever you want to do.
Start with a certain set of
services that already
exist and explore.
Azure is a place of
incredible innovation,
and every month I see
newer updates coming
out in the AI blog,
and it's incredible.
Unlike a lot of other AI
systems and companies,
Azure is like an end
to end ecosystem.
It's not just Cloud.
It's not just AI.
It's an ecosystem. You can build
anything with a set of
services that we have.
Even if you want to
build something custom,
the beauty of Azure is
they allow you to bring
your own data and train on top
of the Azure existing models
called custom models.
It just has everything.
As a start up, you just
have to start fast,
go to market fast.
I think Azure is impressive in
the way it is for
developers for start ups.
The whole ecosystem
is impressive.
We've seen a huge value and
our own developers
are happier than me.
It's not just customer
satisfaction.
It's even developer
satisfaction.
Vitra like I said, today already
translates the top four
or five content
formats in the world.
Going forward, the vision
of Vitra is to translate
literally any type of content.
We want to translate
virtual reality,
augmented reality,
mixed reality.
All of this happening
with just one click.
That's how we portray what
Vitra translate videos images,
AR, VR, etc with just one click.
That's what is the
plan going forward.