Atlanta, GA · AI products and speech models
Oncilla AI builds and operates its own AI products, and trains its own speech recognition and text-to-speech models — starting with Persian. Quiet, practical AI, made to be used every day.
Four AI products, each designed, built and operated by Oncilla AI. None of them is client work — we build them, and we operate them ourselves.
Fronta HealthHealthcare operations
Insurance verification, patient support and scheduling for small clinics.
BrandPulseAI visibility · GEO
AI visibility and GEO: monitors brand mentions across ChatGPT, Claude, Perplexity and similar assistants.
FlashVuData analysis
Natural-language spreadsheet analysis for agencies.
OverlineSpeech · translation
Real-time transcription and translation, as a browser extension and an app.
We trained and fine-tuned our own speech recognition (ASR) and text-to-speech (TTS) models for Persian (Farsi). Our ASR model is fine-tuned from NVIDIA's Nemotron streaming ASR model on 1,181 hours of Persian speech, and it transcribes in real time. Both models are released with open weights.
8.8%
word error rate, compared with 24.2% for NVIDIA's stt_fa baseline
26.0%
word error rate, compared with 54.4% for the same baseline
Built for languages beyond English. Next, we're extending this work to other languages that existing voice AI serves poorly.
Word error rate: lower is better.
Alongside our own products, we take on a small number of custom projects. If your problem sits close to what we already work on — speech, language models and production AI systems — we'd like to hear about it.
Tell us about your project →Oncilla AI is named after the oncilla, a small wild cat known for picking the right target and landing it. We work the same way: a few products, made carefully and run for the long term.
Co-founder · AI researcher & system architect
Mehdi has spent 15+ years designing and deploying machine-learning systems in production, across healthcare, finance, consumer technology and enterprise analytics. His work centers on large language models, NLP, and agentic and retrieval systems, with a focus on evaluation, grounding and reliability.
Earlier, he was a tenure-track Assistant Professor of Computer Science, researching NLP, knowledge graphs and semantic search.
Built: Fronta Health, BrandPulse, FlashVu, Overline, and our Persian speech recognition and text-to-speech models
“Building things that are clear, useful and durable — AI that quietly makes people's work easier.”
Co-founder · Technical strategist & product designer
Parisa pairs a Doctor of Pharmacy background with full-stack development and UI/UX design. Over a decade across pharmaceutical sciences and technology — from medical research at the University of Georgia to building platforms for biomedical firms and startups.
She turns technical requirements into products people find simple to use, and brings the rigor of clinical research to how our products are designed, documented and shipped.
“Combining analytical rigor with creative design to engineer solutions that matter.”
Questions about a product, access to our speech models, partnerships or a custom project — write to us and we'll reply within 24 hours.
We'll reply within 24 hours. Keep an eye on your inbox.