AI R&D that reaches production,
for fast-moving companies
We partner with ambitious companies to solve hard research and engineering problems in AI.
Selected Work
-
Adversarial audio perturbations for privacy
Perturbation algorithms that raised word error rate from 2% to 30% on Deepgram, Assembly AI and GPT-4o, while the audio stayed intelligible to people. For a US seed-stage startup.
-
Whisper fine-tuning and evaluation infrastructure
Whisper fine-tuned for Indian languages to <2% WER, with the evaluation framework that drove model selection and a GPU pipeline subtitling an hour of audio in under a minute.
-
Speech recognition
A TDNN recogniser at <8% WER, deployed for real-time transcription of 911 calls in Motorola CommandCentral — software used in ~60% of US 911 calls.
-
Speaker diarization
An x-vector system trained on ~10,000 hrs, shipped in two Motorola products, outperforming every commercial API tested. Plus its CPU-only production inference system.
FAQ
How does an engagement usually start?
We start with an initial call to understand the problem/requirements. Once the problem/requirements are undetstood, we prepare a written proposal which is then mutually modified and agreed upon. In some cases, before the start of a longer relationship, we do a short 1-2 weeks of engagement to establish mutual fit.
What kind of problems do we work on?
We work on AI R&D problems which are hard to solve. We have the most experience on working in speech and audio problems. However, we have breadth in most areas of AI and generally can quickly adapt to new domains. Some sample problems we have worked on can be found in the work section.
What are the different types of engagements/contracts and how does billing work?
There are 3 different types of contracts that we typically engage in. a) Fixed quote for well defined work scope: In this model, the work scope is well defined and a fixed charge for each milestone is defined. b) Pay by the hour: In this model, the client pays by the number of hours per week, agreed in the contract. c) Retainer: In this model, we allocate a fixed number of hours per month for a monthly retainer.
Who owns the code and other resulting IP?
The code and any other related intellectual property is owned by the client unless explicitly agreed.
Where are you based out of?
We are based out of India and work remotely.
Have a hard problem?
We'd like to hear about it.