TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
David Lequin describes an internal tool that lets employees publish AI-generated HTML dashboards, reporting that weekly use reached about the same level as Looker’s at his company. The report is a first-person account, not an independently verified comparison; it also says the prototype had serious security flaws and does not provide a completed account of the replacement’s security redesign.
David Lequin reports that a small internal service for hosting AI-generated HTML dashboards reached roughly the same weekly user count as Looker at his company, which he says pays about $150,000 a year for the analytics platform. The account describes a fast-built alternative that workers adopted for flexible, custom-looking dashboards, while also acknowledging that its early security protections were inadequate.
Lequin says his company has about 500 employees and a three-person data platform team responsible for infrastructure including Snowflake, dbt and Airflow. An engineer built the first version of the HTML-hosting service in an afternoon in February. It stored pages in S3, used the company’s internal infrastructure and initially placed access behind single sign-on. A small Python publishing script was packaged as a Claude plugin so users could upload pages and share links.
The service began with static HTML files. In March, the team added a live-query API using a viewer’s Snowflake OAuth token. In Lequin’s description, that meant Snowflake’s existing roles determined which data a viewer could access. The team added email and Slack exports in April and parameters in May. He reports weekly viewers rose from 23 in March to 136 in June, and says the tool had as many weekly users as Looker, possibly more, after eight months. These are figures from the report; it does not provide an independent audit or a detailed definition of active use.
Lequin attributes adoption to the ability to generate and refine dashboards as free-form web pages, rather than assembling them from fixed chart tiles. He says workers in finance, operations and marketing used AI tools to create pages, and that executives shared links. The account also says users developed features and scheduled refreshes on their own. Those examples reflect the author’s observations at one company, not evidence that the service can replace enterprise BI across organizations.
Why the HTML Tool Gained Users
The report points to a possible shift in how some employees consume company data: instead of opening a standard dashboard, they can ask an AI assistant to produce a custom page and share it. For a data team, that approach may reduce the time spent building one-off visualizations and give users more control over presentation. It also puts more responsibility on the organization to govern the code, queries and access behind those pages.
The comparison with Looker is about reported weekly usage, not feature parity or verified cost savings. Lequin gives a roughly $150,000 annual Looker figure but does not say the company cancelled the license, calculate the internal service’s full operating cost, or show that it handles every BI task. A home-built tool can complement an established platform without replacing its modeling, administration, support or governance functions.
The security issue is central to the story. Lequin says dashboard JavaScript initially ran on the same origin as users’ sessions, with no dashboard-level permissions, allowing pages to act with the viewer’s access. He characterizes the setup as a major cross-site scripting risk. Adoption, in other words, came before a mature security design; the report does not establish that the service is safe for broad deployment.
data analytics and business intelligence books for data teams
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Static Pages to Live Queries
Lequin’s account begins with a limitation he saw in early “chat with your data” demos: they could produce valid SQL that returned incorrect figures. He says newer AI coding tools became more useful when they could work in the company’s repositories and run queries to check results. The report names Claude Code with Opus 4.5 and describes adoption increasing in early January, but supplies no benchmark or independent evaluation of the model’s accuracy.
The immediate problem was distribution. AI-generated pages saved to personal folders were difficult to share, so an engineer built a service to host HTML and provide internal links. The team’s existing Kubernetes, Terraform and Helm infrastructure helped them deploy it. Adding Snowflake OAuth then connected queries to each viewer’s permissions, according to Lequin. This history matters because the system grew from a sharing utility into a live analytics application, increasing both its usefulness and its security obligations.
“Security, at this point, was bad. Like, really bad.”
— David Lequin, describing the early service
AWS S3 cloud storage and security guide
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Security and Replacement Status
The supplied account ends during the team’s discussion of how to make the tool safer. It says the group wanted to keep the same S3 bucket, avoid migration or disruption, and preserve easy publishing, but does not explain the final security design, whether it was completed, or whether the system underwent a formal review. It also does not clarify whether the early same-origin risk remains in the version currently in use.
Other key details are not established: the exact date of publication, how the company measures weekly users, whether Looker’s own usage declined by a third as the report says, and whether the company reduced or cancelled its Looker contract. The $150,000 figure is presented as an estimate based on an average license, not a disclosed invoice. There is no independent confirmation of the usage figures, cost comparison or claims about dashboard quality.
Snowflake data governance and access control book
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A Safer Publishing Model
The next step described in the report is a redesign of the service’s security model without disrupting existing users or making publishing harder. Lequin says the team began by identifying what it wanted to preserve, including the S3 storage and one-prompt publishing flow. The source material does not give a completion date or describe the safeguards ultimately adopted.
For readers evaluating a similar approach, the unresolved test is whether the organization can retain the flexibility of generated HTML while restricting what a dashboard can access and do. The report offers an account of adoption and a warning about the prototype’s risks, but no final outcome on the rebuild, a decision about Looker, or a verified comparison of total costs.
Claude AI coding and prompt engineering guide
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Did the company replace Looker?
The report does not say that it cancelled or fully replaced Looker. It says the HTML tool reached roughly the same weekly user count, possibly more, while Looker’s usage was reported to be down by a third.
How much did the company pay for Looker?
Lequin estimates the company’s Looker cost at about $150,000 a year, describing it as an average license cost for roughly 300 seats. The report does not provide a contract or invoice verifying that amount.
How did the HTML dashboards access data?
Lequin says the team added a live-query API using each viewer’s Snowflake OAuth token. The stated design relied on Snowflake roles to limit which data a viewer could query.
Was the early system secure?
By Lequin’s account, it was not adequately secured. He says dashboard JavaScript ran on the same origin as the user’s session and that every employee could read every dashboard. The provided material does not describe the completed fix or confirm the current security posture.
How many people used the tool?
Lequin reports 23 weekly viewers in March and 136 in June. He later says the service reached roughly Looker-level weekly use, but the report does not provide a measurement method or independent verification.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
