omerTT
omerTT is the solo developer behind Calibra, an open-source toolkit for robotics dataset observability. Calibra audits dataset integrity, measures quality and behavioral coverage, and selects smaller quality-aware coresets so robotics teams can train policies with less data and GPU compute.
At a Glance
- Robotics teams building imitation-learning policies
- Robot-learning researchers and AI researchers
- Teams collecting LeRobot, Isaac Lab, robomimic, RLDS, or ROS2/MCAP demonstrations
- Teleoperation and robot-data engineering workflows
- +1 more
AI Tools by omerTT
(1)
Calibra
Robot Dataset Audit CLI Tool
Discussions
No discussions yet
Be the first to start a discussion about omerTT
Latest News
Calibra: Cut Robot Training Costs launched on Product Hunt
Calibra v0.8.0 released with measured training-result recording, experiment reports, and simulated/measurement/validation status labels
Calibra v0.7.1 released with Dataset Integrity workflow and expanded validation checks
Calibra v0.7.0 released, making Dataset Integrity the first step before quality, diversity, and coreset selection
Products & Services
A Python CLI/library for robotics imitation-learning dataset observability and coreset selection. It combines integrity checks, quality scoring, behavioral coverage analysis, episode review, quality filtering, and greedy max-coverage pruning; the current package is distributed on PyPI under the name calibra-robotics.
A browser demo where users enter a public LeRobot dataset ID to inspect integrity, quality, and coverage without installing the package.
Commands and reports for comparing full, random, and Calibra-selected datasets across retention levels, recording real experiment outcomes, generating dataset quality cards, and producing partner-facing case-study reports.
Market Position
Calibra positions itself as robotics-specific dataset observability rather than a generic data-labeling or experiment-tracking product: it connects integrity diagnostics and episode-level root causes directly to behavioral coverage and training-set pruning. Its local, open-source workflow and support for robot-learning formats differentiate it from general-purpose dataset tools; its stated benchmark advantage is preserving rare behaviors better than random coreset selection while reducing the amount of training data.
Leadership
Founders
Omer Tahtaci (omerTT / omertt27)
Computer Science/Computer Engineering student at Sabanci University; his public GitHub profile describes him as a CS major, builder, and founder of Calibra and Mergen. He publishes Calibra through the omertt27 GitHub account, the omert27 Hugging Face account, and the PyPI maintainer account omertahtaci.
Executive Team
Omer Tahtaci (omerTT)
Founder and solo developer
Computer Science/Computer Engineering student and builder; maintains the Calibra GitHub repository, Hugging Face Space, package releases, and documentation.
Founding Story
omerTT says Calibra was started to address a gap in robot-learning workflows: teams had strong tools for training policies but lacked a practical way to understand dataset quality before spending GPU time. The initial vision was an open-source, local-first workflow that audits demonstrations, identifies problematic episodes, measures diversity/coverage, and recommends a smaller training set.
Business Model
Revenue Model
The core software is available as a local Python package under the Business Source License 1.1, free for research and internal business use; commercial hosting or managed-service use requires a separate commercial license. The public Hugging Face demo is available without an account.
Pricing Tiers
BSL 1.1 permits research and internal business use; the license states that commercial hosting requires a separate license.
Target Markets
- Robotics teams building imitation-learning policies
- Robot-learning researchers and AI researchers
- Teams collecting LeRobot, Isaac Lab, robomimic, RLDS, or ROS2/MCAP demonstrations
- Teleoperation and robot-data engineering workflows
- Organizations seeking lower training-data and GPU costs
- Pre-training health checks for robot demonstration datasets
- Auditing and repairing data collected for imitation learning
- Selecting smaller training coresets to reduce GPU hours
- Finding rare behaviors and preserving behavioral coverage during pruning
- Continuous-integrity checks in robotics data pipelines and CI
- Teleoperation quality monitoring and real-time data feedback