What shipped, week by week.

The workspace UI is now bilingual. Every top-level page — Devices, Reports, Apps, Spaces, Upload, Tables, Jobs — has a fully translated Japanese version.

Switching languages happens in one click from the user menu. Your preference is persisted, and the agent picks up your locale automatically so its responses match the UI language.

If your team works partly in English and partly in Japanese, the per-user setting means each member sees the interface they’re most comfortable with — no workspace-wide flag needed.

Visitors who land on the wrong language version of tablize.com can now switch with a single click — the language toggle lives in the top nav, visually subordinate to the primary “Try free” CTA.

We start with English and Japanese. Other locales are translated but await editorial review before they’re exposed in the switcher (they’re noindex’d in the meantime to avoid serving stale copy).

The insights capture service that powers our model-bench and regression suites got two reliability improvements:

• Graceful shutdown: SIGTERM now triggers a full drain of in-flight workspace snapshots before exit. Previous behavior could drop the last batch on rolling restarts.
• No more Handle::block_on: Snapshotting workspace state used a blocking call inside the async runtime, which could deadlock under load. Now fully async.

No user-facing change, but our internal bench runs are now reliable enough to gate every PR on regression results.

Some questions deserve a fast, confident answer. Others deserve the agent to slow down, sample the data first, do a sanity check, and show its work. Now you get to pick.

Standard mode (default) is what Tablize has always done — fast, accurate, baseline cost.

Deep Analysis mode triggers a 9-step Verifiable Reasoning Protocol: clarify the question, design the data plan, sample rows, run the query, sanity check, cross-verify, do the math inline, declare assumptions, conclude. Extended thinking is enabled (4K token budget). It costs 2-3× more tokens and adds 50-100% latency, but for decisions that matter you can trust the answer.

The mode is per-turn, not per-session. Start in Standard, hit a hard question, flip to Deep for one message, flip back. The agent re-evaluates each turn with the current setting.

In the UI, Deep Analysis turns are tagged with a verify chip and a gold left border so you can spot them later.

We’ve reshaped how we talk about Tablize.

We used to lead with “Business Agent” — and it scared off the very people we built this for: solo devs, makers, small DTC teams, freelance analysts. “Business Agent” sounded enterprise. Tablize is not enterprise.

The new framing: Tablize is a Data Agent. The analyst seat you never filled. Free to try with your own data. Grows with you from one CSV to a team’s full data stack.

What changed:

• Homepage rebuilt as an editorial vignette of four people each asking their data one question that matters to them today.
• Pricing rewritten as a four-step ladder — Free / Plus / Pro / Max — each tier matched to a specific user state (“just answer my question” → “run our whole data stack”).
• /enterprise page is where the Business Agent narrative lives now, for teams who actually want that layer.
• /use-cases (Templates) and /solutions added to the nav.

The product underneath hasn’t changed. Only the words.