From copy-paste purgatory to one-click tables: automating Bloomberg screenshot data entry

If you’ve ever worked with Bloomberg terminal data, you know the drill. A screenshot lands in your inbox — a PRA (Portfolio Risk Analytics) table, a pricing grid, a yield curve — and now someone has to get those numbers into Excel.

The “old way” looked like this:

1. Select-all on the screenshot data, or manually retype what can’t be selected.
2. Paste into Excel, and watch it land as one garbled column, or as text mashed into the wrong cells.
3. Manually split columns, fix misaligned rows, strip stray characters, re-align headers.
4. Repeat for every single report, every single day.

None of this is hard, exactly — it’s just slow, repetitive, and error-prone. A five-minute task done fifty times a week isn’t a five-minute task anymore; it’s hours of manual data wrangling that produce zero new insight, just friction between “the data exists” and “the data is usable.”

Source

What changed: an AI-assisted internal tool

Working with Claude, we built a small internal tool that closes that gap directly: employees upload (or paste) a screenshot, and get back a clean, structured table ready to export as CSV — no manual column-splitting, no retyping, no Excel gymnastics.

The workflow now is:

1. Paste a screenshot (or open a file).
2. The tool runs OCR and automatically reconstructs rows and columns.
3. Review the result in a live preview, nudge a couple of sliders if a table is unusually dense or oddly spaced.
4. Export straight to CSV, already shaped correctly for Excel.

What used to be a multi-step manual cleanup process is now a paste and a click.

Solution

Tech stack, and why we chose it

• Python + Tkinter — a single-file desktop GUI. No install of a browser app, no server, no IT ticket to whitelist a web tool. Just a script (or a double-clickable .bat launcher) that runs locally on the employee’s machine.
• EasyOCR — handles the actual text recognition. We chose it over a cloud OCR API for two reasons: financial screenshots (portfolio positions, pricing data) shouldn’t need to leave the building to get parsed, and EasyOCR’s recognition model runs fully offline after the first download — so once it’s set up, the tool works with no ongoing dependency on external services, uptime, or API cost.
• A custom grid-reconstruction algorithm, not just raw OCR output. Raw OCR gives you a bag of text boxes with coordinates — it doesn’t know what’s a “row” or a “column.” The real engineering work went into turning that bag of boxes into a table: clustering text into rows by vertical position, then voting across all rows to find genuine column boundaries (rather than naively merging box edges, which collapses dense financial tables into a single column — an early bug we specifically tested for and fixed). This is the difference between “OCR that reads text” and “OCR that understands tables.”
• Tunable parameters exposed as sliders (row tolerance, column gap, confidence threshold) — because no OCR pipeline is perfect on every screenshot, and giving employees a couple of intuitive dials beats sending every edge case back to engineering.

The rationale throughout was: keep it simple, keep it local, keep it fast to run and cheap to maintain. A single-file tool with no server and no cloud OCR bill is easier for one person to own and for anyone else to pick up later.

Why this matters

This isn’t a flashy feature — it’s the unglamorous, high-frequency kind of work that quietly eats hours across a team every week. The value isn’t in the AI doing something impossible; it’s in AI-assisted development making it cheap enough to build a tool that solves a small, specific, recurring pain point that would never have justified a full engineering project on its own. That’s the real productivity unlock: not one big automation, but the ability to knock out dozens of small ones, each returning time to people who’d rather be analyzing data than reformatting it.