Version history of the AI engine that operates this company. The full story is on the engine page; the source materials each step was built from are listed on the methodology page.
Ambient presence. The engine moved from request-driven sessions to a continuous loop: scheduled pulses and watchdog checks run between conversations, and a nightly reconciliation compares what the internal registers claim against the live systems, flagging contradictions instead of letting them accumulate. The loop is event driven as well: incoming email, orders and system alerts start work on their own, without a human prompt.
Multi-model support. Kimi K3 was added alongside the primary models for design work and independent critique: it runs the design pass on our sites and serves as an outside reviewer on quality audits, so the same work gets judged by more than one model family. In our own use it earned the seat: its audit of our rental store found real gaps the primary models had passed over, and its design pass set the current look of our sites. At its July 2026 launch, the vendor's published benchmarks placed it at the frontier for web research and repo-level engineering as an open-weight model at a lower token price. Our own internal numbers stay unpublished until they come with a methodology note.
Governance instruments for an AI board: a published authority model describing what the board may decide alone and what requires the founder, decision records kept in tamper-evident hash-chained logs, independent audit checks, and a publishing gate that holds outbound actions such as social posts, email and deployments until approved.
Measurement and a quality gate. A verdict log rates significant outputs RIGHT, WRONG or OBSOLETE against what actually happened, a weekly retrospective reviews them, and a noise rule treats one data point as noise, not signal. The knowledge base shifted from the founder's personal notes to publicly available university course materials.
Embedded critic. A verifier reads every significant output against the knowledge register before it reaches a decision, and can block or query. The generator's first draft is no longer the final word.
Roles. One AI became a board of six, each with a charter and a defined knowledge domain: product, finance, marketing, security, customer success, operations.
Looking for product release notes instead? Those live in the HyreKit changelog.