# Stan Kirdey

AI engineer working on full-duplex interaction systems at Inflection AI.

## Stan Kirdey · AI engineer

I’m back at Inflection AI building full-duplex voice systems. I want a model to hear you start to interrupt while it’s still speaking and know when to let you finish.

- [Permalink](https://stankirdey.com/about#about-stan)
- [GitHub](https://github.com/skirdey)
- [LinkedIn](https://www.linkedin.com/in/skirdey/)

## Netflix · Finding the lines that explain a failed build

At Netflix, we used distilled transformers to process petabytes of observability data. For an engineer investigating a failure, the useful output could be just a few lines.

For build logs, William High and I used embeddings and locality-sensitive hashing to group log lines by meaning, even when their wording changed. In one build, our report narrowed 6,892 lines to three for an engineer to inspect.

- [Permalink](https://stankirdey.com/about#netflix)
- [See the build-log examples](https://netflixtechblog.medium.com/machine-learning-for-a-better-developer-experience-1e600c69f36c)
- [How we ran inference in the log stream](https://aws.amazon.com/blogs/opensource/how-netflix-uses-deep-java-library-djl-for-distributed-deep-learning-inference-in-real-time/)

## Inflection AI · How do you reward an open-ended answer?

I led post-training and open-source model work at Inflection. We trained Qwen and Llama models with reinforcement learning, using verifiable rewards when an answer could be checked and judgment-based rewards when several answers could be good. Those models achieved our best results on internal evaluations.

In a separate experiment, Andrei Kanavalau and I taught a byte-level language model new languages. We kept its main transformer frozen and retrained just three modules, about 4% of its parameters.

- [Permalink](https://stankirdey.com/about#inflection-ai)
- [Read our multilingual experiment](https://inflection.ai/blog/one-brain-many-tongues-multilingual-blt-with-just-4-new-parameters)

## Hark · Giving computer-use agents a place to practice

On Hark’s founding team, I helped build simulations and reinforcement-learning environments where agents could practice using a browser and operating a computer. We also worked on proactivity, including the question of when an agent should take the initiative.

Hark Handoff later recorded 97.7% average task success on the human-evaluated Online-Mind2Web browser benchmark in August 2026.

- [Permalink](https://stankirdey.com/about#hark)
- [Watch Handoff use a browser](https://hark.com/articles/introducing-hark-handoff)
- [See the human evaluation](https://huggingface.co/spaces/osunlp/Online_Mind2Web_Leaderboard/blob/main/human_Mind2Web-Online%20-%20Leaderboard_data.csv)

## Clark Labs · A personal agent that keeps working when you close the app

I build Clark for tasks you can hand off from your phone. Ask it to compare prices across a few sites, close the Android app, and come back on the web to a report with the prices and links it found. Clark does the work on its own computer. The Android app has reached 1,000+ downloads on Google Play.

I care about making that affordable for everyday use. For developers, our Managed Agent v1 rate card lists $0.75 per million input tokens and $2.50 per million output tokens for an agent with browser and file tools.

- [Permalink](https://stankirdey.com/about#clark-labs)
- [Try Clark on the web](https://www.clarkchat.com/)
- [Get Clark on Android](https://play.google.com/store/apps/details?id=com.clarkslabs.mobile)
- [Managed agent pricing](https://cuarouter.com/v1/agents)

## Amplifier Health

I founded Amplifier Health to study what a voice recording can tell us about health. We built foundational audio models to learn from the acoustic signal itself, including information a transcript leaves out.

- [Permalink](https://stankirdey.com/about#amplifier-health)
- [Read my introduction to the model](https://www.linkedin.com/posts/skirdey_amplifier-health-unlocking-voice-as-a-vital-activity-7219366227930968064-D8J2)

## VoiceRestore

Clean and damaged recordings of the exact same speech are hard to collect. I trained a flow-matching transformer by introducing damage synthetically, so one model could repair noise, reverberation, compression and missing frequencies.

- [Permalink](https://stankirdey.com/about#voicerestore)
- [Hear the before and after](https://github.com/skirdey/voicerestore#example)
- [Read the paper](https://arxiv.org/abs/2501.00794)

## Tenyx

At Tenyx I researched continual learning. The question was how to teach a model a new skill while preserving the things it could already do.

- [Permalink](https://stankirdey.com/about#tenyx)
- [Explore the team’s model and evaluations](https://huggingface.co/tenyx/TenyxChat-7B-v1)
