Engineers and researchers, on one problem.
Apriori is founded and built by IIT Delhi engineers and former Microsoft systems engineers, with research experience spanning large language models, real-time speech, and behavioural simulation. Our team includes SURA research awardees in AI and machine learning, and we publish our benchmarks rather than assert them.
Built by brains from IIT Delhi
Part of immersion programme by
- Antler
Powered by startup programmes
- Microsoft
- Y Combinator Startup School
Programme participation, not investment or endorsement. We are confirming each programme's brand-usage policy.
We are engineers and researchers building voice agents for the hardest audio environment in the world: an Indian mobile call, mid-traffic, mid-sentence, switching languages. We build for the one outcome our customers actually pay for: the parcel gets delivered.
We measured voice AI before we sold it.
Apriori began as a measurement instrument: send us the recordings of a voice agent’s calls, and it would tell you what the callers actually experienced. Response times measured from the audio, not from anybody’s server logs.
Building it taught us where voice AI in India actually breaks. Not on clean English, but on the sentence that starts in Hindi and finishes in English, on a moving auto, at 8kHz.
So we built the agent, for the one problem where an Indian D2C brand loses real money every day: a cash-on-delivery shipment that comes straight back. Two windows decide that outcome, and we work both of them.
The instrument stopped being a product and became our proof. We publish signed measurements of our own agents, from the recordings of your actual calls, using a method that is public in full. Copying that is expensive for a competitor, because copying it means publishing their own bad numbers.
Four rules we would rather lose a deal than break.
Publish the number, or say nothing
There is no figure anywhere on this site that we cannot produce the working for, and every one of them carries its source on the page rather than in a tooltip. Where we do not have the data yet, the site says what the product does rather than what results it gets.
Include the calls we got wrong
Every report we produce contains the calls the agent handled badly, listed individually. A report that only contains good news is not a measurement, and everybody reading it knows that.
Never fake a customer
No invented logos, no written-up case studies of pilots that did not happen, no synthesised recording presented as a real call. A customer is named here only once they have given written permission.
Be specific about the limits
We will tell you which languages are genuinely production ready, which courier write-backs are live today, and which calls our own instrument cannot measure. Vagueness in this market is almost always hiding something.
The team
Names, photographs and links go here once the founders have signed off on how they are described. We would rather leave this section visibly unfinished than fill it with stock photography. Ask us and we will introduce you to the people who would run your deployment.
Where we are
India. We reply to email within one working day, and we are happy to sign an NDA before you send us anything at all.