Data → decisions. From hypothesis to ship-gate.
Open to Data / Product Analyst roles.
I help product teams make decisions on data: A/B tests with reproducible methodology (CUPED, AA-tests, Bonferroni, ship-gates), retention analytics, and SQL dashboards that don't lie.
4+ years at the intersection of product, data, and experimentation. Synthetic data, real methodology.
Start here CV GitHub · LinkedIn · Graph
End-to-end neobank analytics narrative — funnel, A/B test, retention, segmentation. Four sequential projects on synthetic fintech data with reproducible seeded pipelines and ship-gated statistical methodology.
Notjustdashboards—adecisionsystem.Everymetricanswersaproductquestion,everyexperimentpassestheship-gate,anddataspeakswithonevoice.
Calibration-driven A/B testing methods, validated by simulation — SRM, CUPED, delta-method ratio metrics, alpha-spending boundaries, always-valid p-values (mSPRT), bootstrap CIs, HTE, multiple-comparison correction, novelty/primacy and switchback designs. Every module ships with an A/A null check and a power/coverage calibration.
Impact: 15 modules, each calibrated by simulation: Type I error ≈ α, power curves, CUPED: SE shrinks by ~corr(X,Y)², Naive peeking inflates Type I error — Pocock/OBF and mSPRT keep it under control, Delta method gives the correct SE for ratio metrics (CTR, RPC) — the naive per-unit t-test is biased, End-to-end pipeline: SRM → CUPED → delta-method CTR → per-segment ATE + BH → novelty check
Full-stack product analytics platform — a Streamlit dashboard reading live Supabase Postgres through Row Level Security, plus an Edge Function ingest API and a concluded A/B experiment with SQL-computed conversion and a chi-square significance test.
Impact: Full-stack: Streamlit UI + Supabase Postgres + Edge Function ingest (API-key auth), Row Level Security on every table — a user sees only their org's rows, SQL views compute funnel, cohort, MRR, DAU, channel conversion — the dashboard reshapes, it does not aggregate, A/B result in the DB: control 32.1% vs treatment 37.2%, p = 0.0034 (chi-square)
Repo is private — available on request
End-to-end portfolio project — instrument a SaaS product with PostHog, generate realistic traffic, and analyze the events into a funnel, cohort retention, an A/B test, revenue/LTV, and time-to-convert, all in Python.
Impact: Typed event catalog (single source of truth) + PostHog capture/identify/group with PII scrubbing, Feature flag → onboarding A/B variant; A/B analysis with chi-square, uplift, Wilson CI + SRM check, Day-N cohort retention, time-to-convert, revenue/LTV, first-feature → upgrade conversion, The same metrics as SQL (BI / interview reference) + interactive Streamlit dashboard, CI (pytest + ruff) + Docker + render.yaml for one-click deploy
Repo is private — available on request
A self-contained product analytics dashboard on a synthetic SaaS dataset (8,000 users). Covers the AARRR funnel, cohort retention, revenue, and user segmentation.
Impact: 8,000 synthetic users, Jan 2024 – Jun 2025, deterministic seed = 42, AARRR funnel: app_open → signup → activate → start_trial → subscribe with step drop-off, Cohort retention heatmap (signup month × months since signup), Revenue: MRR growth, MRR by plan, ARPU, logo churn, Segments: distributions + conversion + ARPU by segment / channel / country / device
Repo is private — available on request
Streamlit dashboard for AI sales-call analytics — a 4-step funnel (greeting → offer → meeting → qualification), step-1 leak breakdown by type, weighted contact loss, client reactions, engagement by hour and day, bot-quality defects, and drill-down into drop-off transcripts.
Impact: 16,891 synthetic calls, deterministic labeling (script markers), 4-step funnel: greeting → offer → meeting → qualification, Main leak broken down by type (silence / hang-up / refusal) + weighted contact loss, Engagement by hour and weekday — when to call, LLM labeling hook (llm_labels.csv) — the pipeline is reproducible without it
Ten end-to-end SQL case studies on a synthetic product dataset (~183k events), runnable on DuckDB with one command. Funnel, N-day retention, DAU/MAU stickiness, LTV, gaps-and-islands streaks, A/B conversion, and revenue attribution.
Impact: 10 self-contained SQL cases (funnel → attribution), DuckDB — no server, no credentials, one command, Regression tests with deterministic invariants per case, Synthetic deterministic data (seed=42, ~183k events)
RFM analysis of bank clients to segment users by Recency, Frequency, and Monetary metrics. Surfaced key customer groups, optimized marketing campaigns, and improved retention efficiency.
Impact: Identified high-value customer segments, Optimized marketing campaigns, Improved retention efficiency
Cohort retention and LTV analysis on synthetic data — retention curves, churn patterns, and revenue/LTV by acquisition cohort. Python pipeline (pandas + matplotlib/seaborn) plus a Tableau-ready export with CSV and Hyper extract.
Impact: Cohort retention matrix with triangular decay, ARPU / LTV by cohort with proper observation-age caveat, Tableau-ready export (CSV + .hyper extract), Reproducible seeded pipeline (seed=42)
A Telegram bot that automates weekly reporting and integrates with external APIs. Cuts time on routine tasks, ensures timely delivery of key metrics, and improves team communication.
Impact: Cut weekly report prep from 1–2h to automated cron, KPI table + sparklines delivered on schedule, Fallback notifications on source failures, Single consolidated pipeline replacing fragmented sources
A scrollytelling guide for improving spoken English in workplace conversations (A2–B1) — built on Astro 6 with an MDX narrative and D3 visualizations. Content and visuals are decoupled: MDX holds the text, data modules hold section structure and viz props.
Impact: Scrollytelling narrative on Astro 6 + MDX, D3 visualizations (workflow, bars, calendar, checklist), Content and visuals decoupled — trusted data module, frontmatter safely merged on top, Interactive viz panels via IntersectionObserver + dark/light theme + reduced-motion, Deployed to GitHub Pages subpath, Node 22, `astro check` in CI
A personal digital garden and Zettelkasten published as a website with Quartz v4 — linked notes, backlinks, graph view, and atomic evergreen notes instead of chronological blog posts.
Impact: Linked-notes Zettelkasten with backlinks + graph view, Atomic evergreen notes over chronological posts, Quartz v4 — TypeScript-extensible, static output, Obsidian-flavored Markdown → public site
The portfolio you are looking at — built with Astro v5, TypeScript, and Markdown content collections. Static output, dark/light theme, RSS, sitemap, JSON-LD, and a base-path-aware URL helper for GitHub Pages.
Impact: Astro v5 + TypeScript + Markdown content collections, Dark/light theme with no-flash inline script, RSS, sitemap, robots, JSON-LD, OG/Twitter meta, Base-path-aware URLs for GitHub Pages subpath hosting
From hypothesis to ship-gate — the portfolio told as a sequence.
End-to-end neobank analytics narrative — funnel, A/B test, retention, segmentation. Four sequential projects on synthetic fintech data with reproducible seeded pipelines and ship-gated statistical methodology.
Calibration-driven A/B testing methods, validated by simulation — SRM, CUPED, delta-method ratio metrics, alpha-spending boundaries, always-valid p-values (mSPRT), bootstrap CIs, HTE, multiple-comparison correction, novelty/primacy and switchback designs. Every module ships with an A/A null check and a power/coverage calibration.
Full-stack product analytics platform — a Streamlit dashboard reading live Supabase Postgres through Row Level Security, plus an Edge Function ingest API and a concluded A/B experiment with SQL-computed conversion and a chi-square significance test.
End-to-end portfolio project — instrument a SaaS product with PostHog, generate realistic traffic, and analyze the events into a funnel, cohort retention, an A/B test, revenue/LTV, and time-to-convert, all in Python.
A self-contained product analytics dashboard on a synthetic SaaS dataset (8,000 users). Covers the AARRR funnel, cohort retention, revenue, and user segmentation.
Streamlit dashboard for AI sales-call analytics — a 4-step funnel (greeting → offer → meeting → qualification), step-1 leak breakdown by type, weighted contact loss, client reactions, engagement by hour and day, bot-quality defects, and drill-down into drop-off transcripts.
Browser mini-games — including some about analytics: A/B Test (chase p < 0.05), Funnel Drop (catch users through the funnel), plus classic Snake / 2048 / Pong. Works on phone too: swipe, drag, tap.
Concrete engagement formats — for teams that need a solved problem, not "analytics in general".
Review of your experiment process: SRM checks, multiple-comparison correction, CUPED, ship-gates. I will find where you ship statistically-significant but business-insignificant changes.
Cohort curves, LTV decomposition, and a retention dashboard on your data. SQL + Python + BI export.
Heavy-query review, gaps-and-islands, window functions, migration to DuckDB/ClickHouse. Runnable synthetic cases.
RFM or KMeans user segmentation with a per-segment strategy and incremental-revenue estimate.
I work at the intersection of product, data, and experimentation. My background spans customer analytics, A/B testing, dashboard design, and building data pipelines that help teams make faster decisions.
Outside of work I maintain an Obsidian knowledge base on product analytics, experiment with AI-assisted workflows, and write about practical data techniques.
