Product Analyst — impact

Experiments

A/B Testing Methodology Toolkit

Task: Make the method guarantees verifiable, backed by a simulation run rather than by a citation
Result: CUPED cuts variance 17.7% (SE 0.00360 → 0.00327) and the delta method finds a +0.51 pp CTR lift at p < 0.0001 with SRM p = 0.8376 — all 18 modules calibrated by simulation

PythonNumPy / SciPy +2

Analytics

Cohort Analysis Dashboard

Task: Make retention and LTV readable per acquisition cohort rather than as a single average
Result: Python inside Tableau: 4 cohort calculations on a TabPy server and 65 tests; the cohort matrix replaces average retention and shows churn speed and where monetization diverges

Pythonpandas +3

Analytics

RFM Analysis of Bank Clients

Task: Replace the single offer with targeted segments, each with its own retention strategy
Result: Six named segments; the top 20% of clients hold 38.2% of revenue (Gini 0.324), and Champions — 6.4% of the base — drive 12.1% of revenue

PythonSQL +1

Analytics

Churn Prediction — Leakage-Free Retention Model

Task: Score the model so its offline quality is not an artifact of leakage
Result: The top 10% riskiest catches 54.4% of churners at 5.44x lift, but only 1.75x on the at-risk subgroup: 69.7% of the cohort is a near-guaranteed negative by construction.

PythonLightGBM +5

Analytics Real + synthetic

SQL Analytics Case Study

Task: Build cases where every query answers a product question and reproduces with one command
Result: The funnel drops 78% at add-to-cart → checkout (54% is the view→cart step); retention falls from ~19.5% (D1) to ~4% (D30); only 3.5% repeat; on real data 72.4% return

SQLdbt +4

Engineering

ClickHouse — ten SQL cases on 50M rows

Task: I needed a lab where every case is scored against a snapshot and an independent engine, and the optimisation win is measured with the query text held constant
Result: The snapshot and seven DuckDB twins agree, slow and fast SQL are pinned as identical by a test, and the projection clears the ≥10× speed-up threshold the suite asserts on a full scan

ClickHouseDuckDB +5

Analytics

Delivery Control Tower

Task: Build a reproducible loop: ETL, marts, order forecast, bottleneck analysis, testable hypotheses and a SQL layer with measured optimizations
Result: Best WAPE 6.01% at h=7 with seasonal_naive; delivery time barely moves with load (rho = −0.092), yet snacks take 40.0 min against 26.6 min for pizza; the mp channel share grew 5.7 pp

pythonsql +2

Analytics Real data

GMV Drop RCA — root cause on real data

Task: I needed to decompose the drop down to a concrete segment and prove the attribution was complete rather than a convenient slice
Result: UK decile-1 explains 56.2% of the named drop (−£195,218.25); the full attribution reconciles to £0.00 against an overall −10.0% decline

Pythonpandas +6

Analytics

Classifieds liquidity — a measurability audit

Task: I needed to measure how far the proxy drifts from the planted truth per category, and to test whether a paid-visibility decision survives that correction
Result: The proxy reads 0.2413 against a truth of 0.3070, an error of −0.0657 (−21.4%); after correction the worst category misses by 1.61 pp, and paid visibility yields −2.47% contacts per session at p = 0.308

Pythonpandas +5

Experiments

Causal / Uplift — CUPED and Individual Treatment Effects

Task: Implement CUPED and uplift modeling and prove neither is biased where the true answer is known in advance
Result: CUPED cuts the standard error 26% (5.5k users per arm instead of 10k), and a pre-registered break-even threshold of 0.05 ships the campaign to new users only.

PythonLightGBM +5

Experiments

CRM Experiment Lab — Segment-Stage Randomization, GCG and Net Effect

Task: I needed a lab where every CRM experiment stage is checked against data with a known answer and can be made to fail visibly
Result: Segment-stage randomization holds SRM at p=0.266, GCG uplift is +0.9968 [0.9433, 1.0502], and the net effect is +50.5 per 1,000 treated — positive, but the unsubscribe guardrail breaches.

Pythonpandas / NumPy +5

Analytics

SERP Rank Lab

Task: Find out how much revenue ranking for money produces, and what that revenue is paid for in clicks
Result: On a synthetic log: +34.45% revenue per search for −16.84% CTR; the rollout returns −14.26% (p = 0.11) and −16.05% conversion → do not ship

PythonDuckDB +4

Analytics

Sales Calls Analytics Dashboard

Task: Take call breakdown off manual review and name the funnel step that loses the most clients
Result: An answer in a minute: which step loses the client and what to fix; the funnel yields weighted contact loss and a first A/B scenario

PythonStreamlit +1

Analytics

ML Fundamentals — scikit-learn Pipeline

Task: Build preprocessing correctly for modern scikit-learn: no leakage, no reliance on a random seed, reproducible at every pipeline stage
Result: A split reproducible without a seed (Train (8, 5), Test (2, 5)), four synthetic datasets and 21 tests (7 + 9 + 5) that keep the pipeline runnable

Pythonscikit-learn +3

Analytics

RAG Evaluation Harness — retrieval and generation metrics

Task: Build a pipeline with an offline eval harness that separates retrieval from generation metrics and runs without API keys
Result: recall@5 and MRR = 1.00 over 20 questions and 0.00 on an 11-question control set of retrieval misses (MRR@all 0.12) — the metric can fall; correctness 0.37 for the offline baseline (faithfulness ≡1.00 is an extractive invariant); retrieval and generation scored separately

Pythonscikit-learn +4

Product

TaskFlow — PostHog Product Analytics Pipeline

Task: Define events without PII leaks, get them into the analytics tool, and turn raw events into decisions
Result: Onboarding test on a seeded stream: completion 55.9% → 70.8% (+26.6%, χ² = 40.10, p < 0.0001, SRM p = 0.90), and 6.3% of completed signups reach a paid upgrade

PythonPostHog +5

Repo is private — available on request

Product

Product Analytics + A/B on Supabase

Task: Own the multi-tenant setup end to end: auth, per-org isolation, event ingest, and the experiment result computed in the database
Result: A/B concluded on a seeded stream: control 29.6% vs treatment 40.7%, +11.1pp at p ≈ 1.7e-10 (chi-square); metrics computed in the live DB for any client

PythonStreamlit +3

Product

Product Analytics Dashboard (Streamlit)

Task: Build the metrics so they read as one coherent story rather than a set of disconnected charts
Result: Funnel, Retention, Revenue, and Segments read from one consistent base; the same UI became the Supabase project's presentation layer

PythonStreamlit +1

Repo is private — available on request

Engineering

Airflow DAG Platform

Task: Replace the manual export with a reproducible, idempotent pipeline with DQ validation, failure alerts and run metrics
Result: Three DAGs on one shared template and 72 tests without Airflow services or a DB; the full DAG runs in 440 s, 413 s of it in the forecast task

PythonApache Airflow 3 +5

Experiments

Data Analyst Telegram Bot

Task: Check a hypothesis right in Telegram — get a correct result with guardrails from a CSV, without BI and without a notebook
Result: From a CSV — p-value, lift and 95% CI in under 30 seconds, plus a power note and a chart; /export returns a standalone HTML report

Pythonaiogram +6

Repo is private — available on request

Analytics

Python Analytics Playground

Task: Turn the routine into shared modules that teammates can extend without fear of breaking existing behavior
Result: pytest coverage ≥80%; modules stay small and single-purpose, requirements live in the PRD, and the tool extends without breakage

Pythonpandas / NumPy +3

Experiments

Browser Mini-Games — Analytics Arcade

Task: Make a metric's meaning obvious from the action itself, not from an explanation
Result: 10 games: 7 analytics and 3 arcade, one file with no server; Playwright smoke tests catch load failures

SVGJavaScript +2

Engineering

Scrolly English Speaking

Task: Close the gap between grammar and live speech: give learners situations where phrases are used, not recognized
Result: Narrative and visualization in sync: the chart changes as the reader reaches a paragraph; a new scene is just a data file plus MDX

AstroTypeScript +3

Engineering

Digital Garden — EQ in IT

Task: Publish a personal Zettelkasten as a static site where the links between notes are visible to the reader
Result: Links over chronology: backlinks make each note a node, and the graph turns note accumulation into a navigable structure

TypeScriptQuartz v4 +2

Analytics

LinkedIn Content Analytics

Task: Answer not 'how many impressions did I get' but 'which format to write next — and is the difference even established'
Result: 50,101 impressions, 9,440 reach, 1.71% ER over 12.10.2025 — 11.10.2026; three of four comparisons landed inside the noise, and that is recorded as the result

Pythonopenpyxl +3

Repo is private — available on request

Analytics

Promo Margin Lab — what a promo rouble actually earns

Task: Measure promo in margin roubles and choose depth and store set under a budget so that the numbers can be re-checked
Result: Promo returns +68.5 m RUB of revenue and −24.4 m RUB of profit; 97.9% of waves lose money at a breakeven depth of 1.0–19.5% against 25.4% actual, and on one budget the optimiser earns +231 496 ₽ versus +136 648 ₽ for top-15%

Pythonpandas +6

Analytics

Product A/B & Funnel Analytics — Conversion versus Retention

Task: Tie the experiment result to the funnel, cohort retention and unit economics, then issue one management verdict
Result: SHIP: activation → paid +19.59% (p=0.000946) with no guardrail degradation; at an equal step conversion adds +10.00% to LTV versus +8.65% for retention

Pythonpandas / NumPy +5

Engineering Real data

OSS catalog — 47 projects from one source

Task: I needed a catalog where the data comes from one machine-owned source, pages are built from it, and a repeated sync does not overwrite authored edits
Result: 47 project pages and 15 metadata fields are built from one TOML file; a repeated render is idempotent, and 26 tests cover all three record-matching branches

AstroTypeScript +5

Engineering

Matemarketing 2026 Programme — a static replica

Task: Take a third-party interface apart down to measurable quantities and rebuild it with no build step and no backend
Result: The grid matches the original to the pixel: 99 px time column, 420 px hall column, 391 px card, 550 px row, grid top at y = 414; tests catch data and layout drift

HTMLCSS +2

Engineering

This Portfolio Site

Task: Make publishing a project a content change, not a code change: a new Markdown file, then a push
Result: Invalid frontmatter breaks the build, not the deploy; no link hardcodes the base; adding a project needs no code changes

AstroTypeScript +2