Kasra Kyanzadeh

Kasra Kyanzadeh builds AI tools for reflection, reading and structured data extraction. His projects include Lightpage, a notebook that responds to its user's thoughts, and Feedpaper, an open-source experiment in calmer information feeds. He also developed extractGPT and argued for evaluating changes across an entire LLM application pipeline. He publishes and maintains repositories under the handle kasrak.[6][4][7][3]

Kyanzadeh built Lightpage with help from friends. He describes the 2024–2026 project as an exploration of how AI can help people flourish. Users can dictate their thoughts; each morning, the notebook supplies a painting, quotation and essay related to what they have been thinking about. Each Sunday, it summarizes the week in a personal letter. Conversation with the AI is optional: his signed explanation emphasizes “the quiet space to think with yourself.”[2][6]

In Lightpage's signed introduction, Kyanzadeh recalls that a combination of reading, prompts from a trusted thought partner and time to digest them helped him through a crisis years earlier. This supplies personal context for the product's emphasis on reflection and action, rather than accumulating notes for their own sake.[6]

As of September 2026, his personal website says he considers Lightpage finished, continues to maintain it and uses it every day. It also records a brief stint on OpenAI's Codex team in 2026, followed by his decision to step away from building commercial software. He is now learning to make films.[2]

Before Lightpage, Kyanzadeh made Feedpaper to retain what he valued in other people's Twitter posts while addressing a feed he found noisy and addictive. It defaults to yesterday's tweets, keeping the displayed list unchanged until the next day. He called this its biggest improvement for him: “I can just check it once a day”. Related tweets and conversations appear together, while GPT downranks and filters posts according to the reader's interests.[4]

Feedpaper's architecture separates collection from presentation. An Electron app collects tweets from the “For You” and “Following” tabs every 30 minutes and stores them in a local SQLite database. With an OpenAI API key configured, it enriches new tweets by extracting entities and scoring relevance. A Next.js frontend presents the feed.[4]

Kyanzadeh released Feedpaper under the MIT license so others could experiment with presenting information feeds. He explicitly framed it as a prototype rather than an intended Twitter client, citing likely conflicts with Twitter's terms of use. He presented it at the AI | UX meetup documented in Latent Space's April 26, 2023 “It's Time To Build AI | UX”.[4][5][1]

His other 2023 work exposed a different application problem: how to judge whether a cheaper model preserves useful results. In his March 18 essay “Request for Product: Pipeline Replay”, Kyanzadeh described working on extractGPT, which used LLMs to extract structured data from web pages. He wanted to assess a switch from GPT-3 to the more affordable ChatGPT model. He argued that instrumenting only model API calls could miss the effects of preprocessing, linked prompts and postprocessing on the final output.[7]

His replay setup logged relevant parameters and user feedback to a database. A Python script sampled examples with both positive and negative feedback, reran the pipeline and compared the old and new tables. The differences he cared about were concrete: missing expected rows, added spurious rows and changed cell values. His proposed tooling would support experiments across the complete pipeline, including changing models and postprocessing together.[7]

Before these AI experiments, Kyanzadeh worked on product development and engineering leadership. At Airtable from 2015 to 2020, he helped build core features and led a product team making Airtable extensible through APIs, scripts and extensions. At Watershed from 2020 to 2023, he joined as its first employee, helped build the first version of the product and managed the product engineering team.[2]