Michael Royzen

Michael Royzen co-founded Phind with Justin Wei and served as its CEO, building AI search and programming assistance around retrieved context and language-model reasoning. He later founded Standard Signal, where he is CEO. His work has involved consequential choices about when to use external models, when to train specialized models, and how to evaluate whether their answers are correct.[2][3][1]

Royzen and his team focused their search product on programmers in June 2022, around their entry into Y Combinator’s Summer 2022 batch. In his 2023 Latent Space interview, he explained that he expected basic summarization questions to become commoditized. He preferred problems that required reasoning, debugging and solving multiple steps. Phind combined internet search or relevant code from a developer’s codebase with a model capable of using that context to answer the question.[1][2]

The team initially maintained its own search index, but replaced it with Bing after entering YC because the index was too cumbersome. Royzen recalled another concession in February 2023: “our own models aren't good enough”. They adopted OpenAI models, then offered GPT-4 free without a login in April. He described the expense as a deliberate way to show users what the product could do; more capable models also expanded its use from documentation questions to fixing broken code.[1]

Royzen disputed the idea that an AI coding company needed to own the IDE. In 2023, he argued that features such as suggesting diffs could be delivered through an extension, and that the team had not encountered desired VS Code functionality it could not implement. The web interface mattered because software development also includes design and planning outside an editor. His longer-term vision was to help developers move from an idea to working software.[1]

User behavior also shaped the interface. Royzen reported pair-programming conversations reaching 60 or 100 messages, making it difficult to preserve the right information within the model’s context window. Requests to pin important messages led the team to add prioritization controls, allowing users to keep selected information in the conversation context.[1]

Phind returned to training specialized models as open-weight alternatives became more capable. Royzen described the team’s approach as adding substantial training data to CodeLlama rather than changing its architecture. The Phind-CodeLlama-34B-v2 model card documents an additional 1.5 billion tokens of programming problems and solutions, formatted as instruction-answer pairs. Phind reported a full fine-tune without LoRA, using DeepSpeed ZeRO 3 and Flash Attention 2 for 15 hours on 32 A100-80GB GPUs, with a training sequence length of 4,096 tokens.[1][5]

Phind reported 73.8% pass@1 on HumanEval for that v2 release, whose model card also warned that testing had been very limited. Royzen regarded HumanEval as too narrow to capture what programmers actually needed. He described using GPT-4 to score answers to realistic questions and releasing models to users for feedback. The team assessed retrieved context separately from answer correctness: bad context suggested a retrieval problem, while a wrong answer despite good context pointed toward prompting or model limitations. He explicitly said they did not directly measure hallucination in their internal benchmarks.[5][1]

In 2023, Royzen proposed exploring reinforcement learning for correctness. He observed that the model sometimes produced a correct answer but did so inconsistently. His proposed experiment was to generate multiple completions for a problem and reward correct answers, aiming to make an existing capability more reliable.[1]

Royzen also recounted how outside collaborators supported the work. During YC, he asked Paul Graham to invest before showing him the product; Graham agreed and later helped choose the name Phind. Graham introduced the founders to Ron Conway, who connected them with Jensen Huang and Nvidia when they needed GPUs. Royzen credited Nvidia with implementing streaming generation for the T5-based models they then used in Faster Transformer.[1]

By September 18, 2026, Royzen was founder and CEO of Standard Signal, a Y Combinator Spring 2026 company. The company describes an autonomous system that researches the world and executes trades. Its technical thesis is that trading judgment can be learned through reinforcement learning on market outcomes, including learning when to abstain from trading. It says decisions are auditable and hard risk limits are enforced by infrastructure rather than entrusted to the model; its website identifies position sizing, drawdown limits and exposure caps among those constraints.[3][4]

Royzen’s earlier work provides context for his experience with model training and retrieval. Before Phind, he built an iOS visual-recognition app and experimented with question-answering systems while studying at the University of Texas at Austin.[6][1]

Royzen worked on SmartLens in 2017; Diana Drake’s June 2018 Wharton interview described its recent launch as his seventh iOS app. He built its technical implementation around offline recognition. An initial neural network ran on the device’s GPU, then selected additional local networks or cloud recognition according to the object. Royzen said it could recognize more than 17,000 general objects offline, while some downstream functions, including product recognition, used cloud networks. In 2023, he recalled filtering and preprocessing ImageNet 22K and fine-tuning Inception V3 for the app.[6][1]

In 2018, Royzen described SmartLens as a stepping stone and chose UT Austin’s research-oriented Turing Scholars program to deepen his understanding of AI. He had interned at Microsoft Research the previous summer. He also advocated developing an ethical framework for intelligent AI before problems became urgent. Reflecting on SmartLens’s demands alongside school and college applications, he acknowledged: “I overestimated myself”.[6]

During college, invoice-extraction experiments brought Royzen to BERT and Longformer. In summer 2020, Yacine Jernite’s long-form question-answering demo inspired him to experiment with BART fine-tuned on Reddit’s ELI5 questions and answers. He then expanded retrieval beyond Wikipedia. Using remaining SmartLens AWS credits, he recalled filtering Common Crawl to roughly the top 10,000 websites and indexing approximately 350 million pages with Elasticsearch. College demands and operating costs interrupted the project; he resumed it with BigScience’s T0 in fall 2021 and demonstrated the resulting system, Hello, on Hacker News in January 2022.[1]

At the same time, Royzen was writing his senior honors thesis on factuality in abstractive question answering. He said the research influenced Hello’s development because models could not reliably recognize what they did not know. His resulting rule was blunt: “You always have to do a search”. He graduated in 2022 and entered YC soon afterward, choosing the startup over a previously secured job offer.[1][2]