Aragt
Research

Advancing the science of intelligent systems.

We work on the hard, foundational questions of modern AI,the algorithms, models, and evaluation methods that will define what these systems can do.

Our approach

Our research agenda is chosen around a simple question: what work, if we did it well, would meaningfully change what AI systems are capable of,and how safely they can be deployed?

We invest in long-horizon foundational research and short-horizon applied work in parallel, and we publish and open-source throughout. The best research doesn't sit behind a paywall.

Areas

Twelve areas we invest in.

From foundation models to alignment, from robotics to African language AI,the surface area of modern AI is wide, and we work across it.

  • R.01

    Foundation Models

    Training frontier language and multimodal models with strong reasoning, coding, and multilingual capability.

    • Pretraining data & curation
    • Scaling laws & efficiency
    • Post-training & preference learning
  • R.02

    AI Agents

    Building systems that can plan, use tools, and complete work reliably over long horizons.

    • Tool use & code execution
    • Planning & memory
    • Evaluation for long-running tasks
  • R.03

    Multimodal AI

    Unified models that reason across text, images, audio, video, and structured data.

    • Vision-language alignment
    • Speech understanding
    • Document & UI understanding
  • R.04

    Computer Vision

    Perception systems for real-world environments,from satellite imagery to medical scans.

    • Geospatial AI
    • Medical imaging
    • Video understanding
  • R.05

    Speech & Audio

    ASR, TTS, and audio understanding for high- and low-resource languages.

    • Multilingual ASR
    • Expressive speech synthesis
    • On-device inference
  • R.06

    Robotics

    Learning-based policies for manipulation, mobility, and embodied intelligence.

    • Sim-to-real transfer
    • Multi-task policies
    • Foundation models for control
  • R.07

    Reinforcement Learning

    Learning from feedback,from human preferences to verifiable rewards in code and math.

    • RL from AI feedback
    • Verifier-guided training
    • Exploration & credit assignment
  • R.08

    AI Safety

    Making powerful AI systems predictable, controllable, and trustworthy.

    • Model evaluations
    • Red-teaming & robustness
    • Oversight of autonomous systems
  • R.09

    AI Alignment

    The science of building models that reliably do what people actually want.

    • Scalable oversight
    • Interpretability
    • Value learning
  • R.10

    ML Systems

    The training, serving, and evaluation infrastructure that makes frontier research possible.

    • Distributed training
    • Low-latency inference
    • Reliable evaluation pipelines
  • R.11

    African Language AI

    Foundation models, datasets, and benchmarks for the languages spoken across Africa.

    • Multilingual pretraining
    • Community-sourced data
    • Language-preserving evaluation
  • R.12

    Scientific AI

    AI as a tool for scientific discovery,in health, agriculture, climate, and beyond.

    • Protein & molecular modeling
    • Climate & earth systems
    • Automated experimentation
Publications

We share what we learn.

Papers, technical reports, model cards, and evaluations,published as we go.

  1. Coming soon · Language
    A foundation model for African languages,technical report
    Preprint →
  2. Coming soon · Agents
    Verifier-guided reinforcement learning for code
    Preprint →
  3. Coming soon · Safety
    Evaluations for long-horizon autonomous agents
    Preprint →