Founding Research Engineer - World's Audio Data

2 нед. назад

80k–180k USD / yearUSASeniorRemote

data collectionmachine learningaudio processing

Founding role focused on creative data gathering and AI model training for regional language voice technologies.

Условия

A mission worthy of being your life’s work.
True ownership of work.
Top-tier teammates that are ambitious, smart and fun.
Top percentile equity and pay that truly value your talent and opportunity cost.

Другое

You will find creative ways to gather the world’s audio and make it usable for training. Beyond scraping the internet, you might buy data stored on people’s phones, give people microphones to wear all day, or install microphones at kiosks around the world—all at 100× less cost than the market norm.
This is a founding role, so your scope will expand and change as UpliftAI grows.
We build voice models and voice applications for regional languages — Urdu, Bengali, Pashto, Greek, and more — so people around the world can use technology by simply speaking.
We gather our own data, build our own labeling tools, train our own models, and design our own voice infrastructure. This fully integrated approach is what lets us deliver the highest-quality models, and the best developer experience, for regional languages.
Over 60% of the world's population only speaks a regional language. Today voice models aren’t good enough in their languages. Large tech companies are unable to give this work tier1 priority because they are busy defending their primary business. In fact, in all big labs regional languages are funded by Marketing/PR budgets. If UpliftAI doesn’t do its work, a large part of the world may continue to experience significant AI technical lag — further worsening the income divide between countries.
But also, it’s very cool making machine speak soooo well (check out our homepage).
Hammad left Apple and took a year off to figure out what to do in the next phase of his career. During this time, he started a side project gathering data for Pakistani languages so people could talk to ChatGPT by speaking (this was before they supported voice mode).
To gather this data, he traveled to factories and rural areas in Pakistan. While speaking with business owners there, he learned that many businesses are not using digital tools (like HR software or inventory management) simply because their workforce cannot read or write. (Btw, Pakistan is the 5th most populous country in the world, but 42% of all adults there cannot read).
That is when the idea first emerged: what if everyone could use technology by speaking? Voice interfaces can finally digitizing businesses that currently cannot. This would increase worker productivity — which at scale lifts the entire GDP of developing countries. Globally, 1 billion people cannot read, and voice interfaces will finally give them access to digital services… often for the first time.
Proof of exceptional talent through past accomplishments. Ask yourself: “What have I done that none of my peers did?”
The ability to move 3× faster than your peers through extreme resourcefulness. You are comfortable solving problems with code, partnerships, fieldwork, or whatever else works.
Strong software and data engineering skills. You can build scrapers, pipelines, and tools that process large amounts of audio.
The judgment to determine which audio will actually improve a model—not merely collect large amounts of it.
Hammad - CEO: ex Apple, ex Amazon, Cornell alumnus. Previously developed voice technology at Apple Siri and Amazon Alexa for about 9 years. I created my first website at age 13 to sell wallpapers & software. This was year 2003 before even wordpress existed. It was hosted over dialup from home. Zaid - CTO: Zaid previously designed and launched an entire AWS service: Bedrock Guardrails. Before that, at Amazon, Zaid wrote a paper on how to optimally plan and pack products in containers to minimize transportation and destination distribution cost in Amazon’s global import supply chain. This work led to $100m/year saving.
Below is one possible plan. We expect you to improve it once you understand the problem.
First 30 days: Understand which data most improves our models, audit how we currently acquire and prepare it, and ship your first new source of training-ready audio.
First 60 days: Launch several unconventional data-acquisition experiments and build the pipelines needed to filter, segment, transcribe, deduplicate, and score the audio automatically.
First 90 days: Prove at least one method that produces high-quality training data at 100× less than the market norm, then begin scaling it across multiple languages.
First year: Build a repeatable system that lets Uplift gather the data needed for a production-grade model in any language within weeks instead of multiple months.