CrestaGrowth

Cresta was founded in 2017 by Stanford PhD students Zayd Enam and Tim (Tianlin) Shi, with their adviser Sebastian Thrun as co-founder and chairman, as a spin-out of the Stanford Artificial Intelligence Laboratory. Enam's doctoral work applied natural-language processing to repetitive knowledge work; an early model recommended behaviours to low-rated Udacity graders based on those of high-rated ones, and Thrun steered the project toward sales and support conversations. The company left stealth on February 3, 2020 with a real-time agent-assist product: models trained on a client's historical chat transcripts learn the phrasing used by top-performing agents and suggest responses to other agents during live conversations. Cresta engineering posts from January 2021 describe a GPT-2 model trained on annotated chats to generate suggestions and a multi-head BERT model for intent classification.

Cresta's granted US patents, all assigned to Cresta Intelligence Inc. and each naming Shi as an inventor, cover that first product: identifying a behavioural target during a monitored conversation (US 11,115,531, priority March 2020), selecting effective phrases to present in real time (US 11,157,695, priority July 2020), choosing voice transformations for an agent's speech by machine learning (US 11,120,812), selecting portions of an audio stream for contemporaneous analysis (US 11,282,507, priority October 2020) and initiating system actions from conversational content (US 11,417,337, 2021).

The company moved to large language models in 2023. On March 27, 2023 it announced generative-AI additions to its platform, and on June 20, 2023 it introduced Ocean-1, which it described as a contact-center foundation model built by combining base foundation models such as GPT-4 with contact-center domain data, using self-instruct and reinforcement learning from human feedback, with the stated goals of better instruction following and lower serving latency and cost. By December 2023 the company said its Ocean models were 7-billion-parameter Mistral and Mixtral fine-tunes served with LoRA adapters through Fireworks AI, and claimed roughly 100 times lower cost than GPT-4 for retrieval-augmented Knowledge Assist. An April 29, 2024 engineering post by Motoki Wu, Ethan Jiang, Chuan Wang and Shi states that Mistral-7B fine-tunes power Knowledge Assist, call summarisation and chat suggestions alongside partner models from OpenAI and Anthropic, evaluated on coverage, contradiction and groundedness checks. On November 7, 2023 Cresta released Opera, a no-code interface for supervisors to define LLM-based rules for coaching and quality assurance. On October 29, 2024 it announced Cresta Voice Virtual Agent, described as a hybrid architecture combining LLM generation with deterministic controls, and on June 9, 2025 an omnichannel AI Agent that carries context between voice and digital channels.

Cresta's applied-research group published FECT (KDD 2025 workshop on evaluation of agentic and generative AI; arXiv, July 2025), a benchmark and annotation scheme for judging the factuality of interpretive LLM claims about contact-center transcripts, with code and data on GitHub. No other peer-reviewed paper with a Cresta affiliation was found. Google Patents lists five US patents and applications assigned to Cresta Intelligence Inc., all filed 2020-2021; no filings covering the Ocean models or the AI Agent were found. The legal entity is Cresta Intelligence, Inc., headquartered in Sunnyvale, California. No acquisitions by Cresta were found.