Aionic Labs develops time series language models (TSLMs) that enable AI systems to read raw signal data directly and reason over it in plain language. Its open-source model, OpenTSLM, scores 99.71 on the TSQA reasoning benchmark—far surpassing GPT-4o—and supports healthcare, industrial, and energy applications. The company is recruiting enterprise partners for a pilot program to deploy its platform across data-rich domains.
- Artificial Intelligence
- AI Agents
- Data & Analytics
- Enterprise Software
- Software Only
Funding
Funding not disclosed
Founders
Product
Problem
Traditional time series analysis tools require manual configuration, expert interpretation, and often reduce complex signals to dashboards or summaries that miss subtle drifts, correlations, and anomalies. This makes it difficult for organizations to detect early warning signs, understand root causes, and decide on timely interventions across operational, clinical, and industrial data streams.
Solution
Aionic Labs builds a horizontal TSLM platform that treats time series as a native modality, allowing language models to read raw multivariate signals directly alongside text. The platform ingests, trains, evaluates, and deploys models capable of open-ended reasoning, anomaly detection, forecasting, and intervention recommendations—all expressed in natural language with confidence scores. Its first model, OpenTSLM, accepted at ICML 2026, uses a Flamingo-style cross-attention architecture to scale to long and multiple time series with near-constant memory usage. The system surfaces findings in plain language, describes how signals evolve, and recommends actions with expected outcomes so domain experts can decide with full context.
Target Audience
Primary customers are enterprises in healthcare, finance, industrials, energy, and telecommunications looking to deploy TSLM-based predictive maintenance, anomaly detection, forecasting, and optimization applications.
Features
- OpenTSLM architecture treats raw time series as a native input modality, avoiding tokenization into text or image plots
- Flamingo-based cross-attention design scales to multiple long series with near-constant VRAM usage, unlike prior quadratic-scaling methods
- Scores 99.71 on TSQA reasoning benchmark (vs. 59.24 for GPT-4o) and outperforms image-, token-, and GPT-4o-based baselines across ECG QA, sleep staging, and activity recognition
- Validated with Stanford Hospital cardiologists: 97% of 84 reviewed clinical rationales rated correct or partially correct using ACC/AHA-based rubric
- Released open source under MIT license; supports text-plus-signal multimodal prompting, multi-series comparison, and forecasting without task-specific fine-tuning
- Horizontal platform covers ingestion, training, evaluation, and deployment with licensing, API, and on-prem options