Situation
PMs and analysts often have no BI access: checking an A/B hypothesis waits for a dashboard, and the exported CSV sits there without statistics. The hypothesis is written and the data is exported — yet p-value, lift and an SRM check still demand a notebook, an environment and someone who remembers the formula.
Task
I needed to bring hypothesis checking into Telegram: get a correct result with guardrails from a CSV — no BI, no notebook, no manual statistics.
Actions
Stack
aiogram 3 + pandas, scipy, statsmodels. The UI follows the Telegram locale (EN/RU); analytics output stays English.
Commands
| Command |
What it does |
/start, /help |
Show commands |
/describe |
Shape, dtypes, missing %, numeric describe, top categoricals, datetime range |
/funnel |
Step-to-step and overall conversion (warns on non-monotonic steps) + chart |
/abtest |
Welch t-test (numeric) or chi-square (binary), p-value, lift, 95% CI, SRM and power guardrails + chart |
/cohort |
Weekly cohort retention per first-activity week + heatmap |
/rfm |
RFM segmentation (recency/frequency/monetary) per user + chart |
/trend |
Time-series trend (daily/weekly/monthly) + chart |
/correlate |
Top pairwise Pearson correlations + heatmap |
/history |
Your last 10 analyses |
/export |
Last analysis as a standalone HTML report with a chart |
/cancel |
Reset waiting state |
| Command |
Columns |
/funnel |
step,users — rows ordered top to bottom of the funnel |
/abtest |
group,metric — metric numeric or binary 0/1 |
/cohort |
user_id,date — one row per activity; date is also accepted as dt, day, ts, timestamp, created, updated, дата |
/trend |
date,metric — one row per observation |
/rfm |
user_id,date,amount — one row per transaction |
/correlate |
any CSV with two or more numeric columns |
/describe |
any CSV |
Sample datasets live in data/samples/.
Engineering
Analytics lives in src/analytics/ as pure, tested modules; CI runs pytest, ruff and mypy. Uploads are capped before parsing — 10 MB / 100k rows (free tier: 5 MB, 10 analyses a day). History is SQLite, the last 10 analyses.
Result
From a CSV — p-value, lift and 95% CI in under 30 seconds, plus a power note and a chart. /export returns the last analysis as a standalone HTML report you can attach to a hypothesis discussion.
Limitations
The bot computes but does not design an experiment: observation independence, novelty and multiple comparisons stay with the analyst. Data arrives only as a CSV — there is no warehouse connection. The free tier is capped at 5 MB and 10 analyses a day. The repository is private — there is no public code, and the card does not link an empty repository.
Documentation
- The repository is private — available on request. No public code; the card does not link an empty/stub repository.
Frequently asked
Which commands does the bot have?
12 commands: /start, /help, /describe, /funnel, /abtest, /cohort, /rfm, /trend, /correlate, /history, /export, /cancel.
What do I send for /abtest?
A CSV with group and metric columns; metric is numeric or binary 0/1.
What does the bot compute besides p-value?
Welch t-test for a numeric metric or chi-square for a binary one, then lift, 95% CI, SRM and power guardrails, and a chart.
Where is the data stored?
Locally only: the last 10 analyses in SQLite. The token is read from the BOT_TOKEN environment variable, .env is gitignored.
How is it deployed?
Long polling needs an always-on process: a Fly.io worker built from the Dockerfile or a local docker run. CI runs pytest + ruff + mypy.
