Skip to content

Kareem Nassar

AI systems engineer • real-time speech, world models, RL/evals

Now: building Candor, a terminal-native way to collect human signal for agent evals, RL workflows, and product feedback loops.

Previously co-founded EzDubs, real-time speech-to-speech translation acquired by Cisco/Webex. Before that, I built ASR and voice systems across startups and large-scale deployments.

Lately I’m thinking about language world models, AR + diffusion architectures, generated-state feedback loops, and what makes agent rollouts useful.

World Models RL/Evals Human Signal ASR S2ST PyTorch

Current notes

  • World models for agents: Qwen-AgentWorld, simulator fidelity, LLM-as-judge rewards, and sim-to-real transfer for agent rollouts.
  • Multimodal architectures: two-tower AR + diffusion systems like Cosmos 3 / Nemotron TwoTower, especially feedback from generated states back into reasoning.
  • Human signal: how to collect better eval, ranking, and preference data close to the agent workflow.

Research Notes

Small public experiments and explanations where the mechanism is inspectable: speech systems, multimodal models, biological learning rules, sparse expansion, and agent eval loops.

Candor — Human Signal for Agent Workflows

2026 • RL/evals • human preference data • terminal workflows

Candor collects human feedback where AI builders already work: the terminal. It supports pairwise comparisons, ratings, labeling, product feedback, and AI-moderated interviews, then returns structured signal for evals, training data, and agent workflows.

I’m especially interested in where human signal fits into agent RL loops: judge calibration, high-disagreement cases, preference data, and evaluation rubrics that survive contact with real users.

Writing

Most short-form writing is on X: world models, multimodal architectures, real-time speech systems, eval loops, and the human signal that makes those loops useful.

EzDubs on Ray-Ban Meta Glasses

Real-time speech translation • wearable UX

A quick demo of Paddy and me talking through EzDubs using Ray-Ban Meta glasses. This is the kind of product surface I like: AI that disappears into a live interaction instead of becoming another dashboard.

Acquisition: Cisco announcementTechCrunch coverage

More EzDubs Demos

Real-time translation • product demos

A couple more demos from the same product arc: real-time voice translation moving from model work into live user-facing surfaces.

RL, Rollouts, and World Models

SAC • replay • rollouts • PyTorch

Implemented Soft Actor-Critic with an AtariNet-style convolutional encoder. Training leveraged replay, MC rollouts, and importance sampling. I’m still interested in the same core problem today: how agents learn from simulated experience, how those rollouts are scored, and when simulated feedback transfers to real environments.

Training code: gist

Smartmiq Desktop Assistant

2021 • PyTorch (Transducer), macOS UI

Spotlight‑style voice assistant for desktop. Trigger actions hands‑free (search, reply to Slack, etc.). I trained my own ASR based on a transducer architecture for tighter control over latency and accuracy.

What I learned: voice on laptops/desktops is useful but often slower than typing in public; UX must bias toward zero‑friction hybrid input (keyboard + voice), and ASR needs strong endpointing.

Smartmiq demo: voice search triggering Google results

Replying to Slack by voice:

Smartmiq demo: replying to Slack messages with voice

Noise Suppression Neural Network

2020 • VoiceFilter‑style masking

Learned a spectral mask to suppress non‑speech noise, inspired by Google’s VoiceFilter. The model improves intelligibility in common home/office environments.

Rover — Music Visualizer via GAN

Rust • DCGAN latent exploration

Mapped audio frequency features to a smoothed latent walk through a pretrained art‑DCGAN to generate synchronized visuals in real time.

  • Compute frequency components per frame
  • Integrate with previous latent for temporal smoothness
  • Render GAN output to framebuffer

Code: github.com/kareemn/rover

Meeting Voice Assistant (Zoom/Meet/Webex)

2018 • Workfit/Voicera/Voicea • Kubernetes

Before high‑quality meeting audio APIs existed, I emulated webcam/mic drivers in a Dockerized headless Chrome to capture >16kHz audio for wake‑word and ASR, then scaled it across a Kubernetes cluster. We also used a virtual webcam to demo the bot live, a surprisingly effective growth loop.

Poynt OS & Platform Integration

2015-2017 • Android OS fork • kernel/HAL • OTA updates • PCI PTS

Led OS and platform integration for Poynt's Android-based payment terminals: Android OS fork work, kernel/HAL integration, OTA update framework, the Poynt App Store, and OS security changes for PCI PTS certification. This was building and shipping an operating system for real payment hardware, not app-layer product work.

The Nay-Nay video below was a stress test of the remote OS/app update pipeline: we pushed a music video to a fleet of test devices at once.