Adam Wiggins

Adam Wiggins is a cofounder of Heroku and Ink & Switch, a software researcher, and a technical writer. His recent AI work includes experiments in personalized email triage and writing about how applications can make language-model results editable, verifiable, and manageable. At Elicit in 2024, he described his contribution as internal journalism: learning from colleagues and using writing to explain their expertise.[1][4][5][3]

In January 2026, Wiggins announced a three-month investigation as researcher-in-residence at Ink & Switch into a personal algorithm for message triage. The goal was to learn which messages mattered to an individual through feedback, rather than requiring manually maintained filtering rules. He framed the project as research with uncertain feasibility, confronting small personal datasets, fragmented communication channels, and restricted APIs.[4]

In his email-classifier experiment, Wiggins preferred separating semantic understanding from preference learning: compute an embedding for each email, then classify it with logistic regression. He hand-labeled Fastmail messages as priority, fyi, or ignore, and added features such as recipient placement and previous correspondence with the sender. The browser prototype used TypeScript, React, and IndexedDB through Dexie, while fetching messages from Fastmail and obtaining embeddings from OpenAI.[5]

Wiggins reported 74 correct classifications out of 90 evaluation emails. With manually selected confidence thresholds—a probability of at least 0.5 and a margin of at least 0.2—the classifier made predictions for 69 emails and got 64 right. The higher accuracy therefore depended on leaving 21 cases unclassified. These results came from one personal dataset, with a 75% training and 25% evaluation split; practical production usefulness remained unresolved. After comparisons with prompted GPT-5.2 and fine-tuned gpt-4.1-mini, he preferred the classifier's simpler, faster retraining.[5]

Wiggins's May 22, 2024 essay, Living Documents as a UX Pattern in AI, argues for integrating AI into richer interfaces: “But a greater opportunity exists: to weave AI capabilities into more sophisticated UIs.” He explains Elicit's editable research tables, where many independent model extractions become cells that users can inspect and revise. Fifty papers and twenty extraction columns produce 1,000 cells, each potentially requiring multiple model calls. The team's interface practices include incremental results, quiet loading placeholders, low-confidence flags, supporting source passages, retries for individual cells, and visible operation costs.[3]

In the June 2024 Latent Space interview with James Brady, Wiggins argued that developers building end-user AI applications need both product development and distributed systems skills to create stable interfaces around unpredictable language models. His role in the accompanying hiring article was to support Brady, its primary author, and surface the team's expertise through writing.[1]

Before his Elicit writing and email-triage research, Wiggins worked on Muse from 2019 to 2023. That period provides earlier examples of his work on software interfaces and product communication. Muse Software formed in summer 2019 as an Ink & Switch spinout with Julia Roggatz and Mark McGranaghan. In his February 2024 retrospective, Wiggins credits Roggatz with the initial iPad prototype and McGranaghan with the product vision and name. His own choice was explicit: “I decided to make my focus for this venture be storytelling.” He worked on positioning, brand, newsletters, and distribution, testing homepage messages before choosing “tool for thought.”[6]

In April 2019, Wiggins coauthored Local-first software: You own your data, in spite of the cloud with Martin Kleppmann, Peter van Hardenberg, and Mark McGranaghan. Their proposal treated copies on users' devices as primary and server copies as secondary, seeking collaboration without sacrificing offline access or user control. They investigated conflict-free replicated data types, or CRDTs, as an enabling technology. The essay's seven ideals covered responsiveness, multiple devices, offline work, collaboration, longevity, privacy, and user control; its prototype evaluation was exploratory rather than formal.[7]

Muse later adopted bespoke CRDT synchronization after experiments with iCloud and Firebase. Wiggins credits McGranaghan with building the sync system and Adam Wulf with the Swift client implementation and persistence migration; Roggatz and Lennart Ziburski developed the Mac experience. In his retrospective, Wiggins describes a 2022 move toward team collaboration and business customers, followed by adoption friction and financing difficulties. By 2023, the team collectively concluded that its approach had reached a dead end. Wiggins stepped away alongside most of the team that fall.[6]

A decade before his AI writing, Wiggins was explaining Heroku's application architecture. His June 20, 2011 process-model announcement describes the Procfile as a way to declare named process types through command-line commands. Web and worker processes could run in isolated instances and scale independently, while one-off commands ran against deployed application code. In a contemporaneous InfoQ interview, he expressed a preference for established Unix tools: “The beauty of Unix: you can achieve quite a lot with a combination of a few sharp tools.”[9][11]

Wiggins wrote The Twelve-Factor App, dated 2011 on his personal site and last updated in 2017 according to the methodology's footer. It synthesizes contributors' experience deploying software-as-a-service applications at Heroku into recommendations for portable applications: explicit dependencies, environment-based configuration, stateless processes, separated build and run stages, development/production parity, and logs treated as event streams.[2][8]

As Heroku cofounder and CTO, Wiggins publicly addressed routing and performance complaints in an April 3, 2013 FAQ. He acknowledged that “we slipped on stewardship of our existing product,” citing stale Bamboo documentation, inaccurate queue-time reporting during 2011–2013, and delayed responses to customer complaints. His account explained the platform's trade-off favoring availability and stateless scaling, and outlined concurrent backends, migration assistance, and improved performance visibility as responses.[10]