# Kultrix time-tracking dataset (anonymised)

Time records from a small software product studio in Lviv, Ukraine, March 2025 to October 2026.
4,271 records, 24,857 hours, 33 client and internal projects, 24 people, 569 tasks (290 with an estimate).
Licence: CC BY 4.0. Summary analysis: https://kultrix.com/research/where-development-hours-go

## Files

- records.csv: one row per time entry.
  - record, date (day the work was logged for), project (P1..), billing (fixed or hourly), developer (D1.., numbered at random within each project), task (T1.., empty if the entry has no task), hours.
  - stage: dev, qa, bugs, design, meetings or other. Assigned by keyword rules on the task name, then the description; the rules are in the summary analysis.
  - entry_method: timer (desktop tracker) or manual.
  - llm_access: whether the person had company-provided Claude Code access on that date. Access was rolled out person by person from September 2025.
  - description: rewritten by an LLM into generic engineering terms; names, business domain, client-specific feature names and third-party services removed.
- tasks.csv: task, project, estimated_hours (empty = no estimate), logged_hours (sum of records).

## Notes

- Project ids were assigned in random order. Developer ids are per project, so the same person has unrelated ids in different projects.
- Removed: names, client and product names, financial data, original descriptions.
- llm_access comes from the date access was handed out in the company chat. Earlier personal use of other AI tools is not recorded.
- Estimates were entered by the team or project manager when the task was created; some were revised later and only the final value is kept.
- Some people logged a week or more of work as a single entry on one date (22 entries exceed 24 hours, the largest is 160). Treat hours per date as hours per entry, not per calendar day.
- The founder's own entries and projects with fewer than 5 entries were left out.