Nikita Fedorov

I enjoy understanding how complex systems behave, documenting what I learn, and sometimes building them.

From geological formations to digital products, I like studying the patterns that shape our world.

Every system leaves traces. Here I collect some of them.

now

Running product analytics across several consumer products, and rebuilding how the team measures whether its own work pays for itself.

Writing up the Monte Carlo habit — why a single number is usually a confident guess wearing a suit.

Slowly turning a decade of private notes into something worth reading in public.

what I’m doing now →

recent traces

Notes in progress

Notes grow in stages — carbon, graphite, diamond — and most never make it past the first. What’s here is what has survived so far.

The first notes are still forming.

all notes →

patterns

What keeps showing up

Different domains, same shapes. Four things I keep running into, whether the system is a reservoir or a product.

Every point estimate is a distribution in disguise

A forecast, a reserve estimate, a delivery date — each one is a single number standing in for a range nobody measured. The number is easy to defend and impossible to calibrate. Ask what it would take to be wrong, and the range shows up on its own.

Activity is easy to count, value is not

Reach, ticket volume, messages sent — systems drift toward measuring whatever is cheapest to measure. The expensive question is whether any of it paid back. Most of my work is dragging measurement from the first kind to the second.

A system isn’t understood until someone can act without you

If the analysis only works when the analyst is in the room, what got built is a dependency, not an understanding. The finished version is the one the business can run alone — a tool, a readout, a rule of thumb that holds.

Teams behave like systems too

Same rules apply: feedback loops, load-bearing parts, failure modes that show up long before anyone names them. Workload patterns predict burnout months ahead of a resignation letter, if anyone is reading them.

systems i’ve worked on

Four systems, four kinds of trouble

Where the patterns above met an actual business. Each card is one system, what was wrong inside it, and what changed. Open a card for the long version; the numbers are the ones I can share.

profitability

The audit that reset acquisition economics

The product's first ROAS-based evaluation of marketing spend — secured under an executive mandate during a P&L downturn, and now run by the business itself.

ROAS framework60% of deals exposedself-serve tooling

full story

Context

A consumer product was buying media on first-deposit volume and "brand reach," with no measure of whether spend actually paid back. When product P&L turned, I secured an executive mandate to audit acquisition and retention end-to-end.

What I did

Rebuilt evaluation around ROAS at the deal and channel level, split budget into performance vs. brand for spend transparency, and redesigned the team's KPIs around return and successfully-deployed budget instead of activity volume.

How

Shipped a self-serve deal-scoring tool that reads out payback at the 2-week and 1-month marks — the team decides who to keep and who to cut without waiting on analytics.

Signal

Roughly 60% of legacy deals didn't survive the math; monthly ad spend was roughly halved with no revenue loss. For the first time in the product's history, profitability — not reach — drives media decisions.

ai tooling

Code review in under a minute

A Claude-based reviewer inside GitLab that checks code against business intent — paired with a governance model that decides who gets AI at all.

< 1 min review latencyintent vs implementationtiered AI access

full story

Context

Team leads were the review bottleneck: every merge request queued behind their operational load, and review depth varied with how busy the week was.

What I did

Built and deployed a Claude-based reviewer in GitLab that validates SQL and Python logic against the linked Jira ticket — checking that the implementation matches business intent, not just syntax — and flags security issues, performance problems, and code-style deviations.

How

Alongside the tool, wrote the adoption policy: AI goes first to analysts who can tell when a model is hallucinating, not to everyone at once. Access is tiered by a user's ability to verify the output — a risk-aware rollout instead of a blanket one.

Signal

First review lands in under a minute, and team leads got back roughly 5–7 hours a week. After adoption by the platform team, the reviewer is scaling org-wide toward 100+ analysts.

experimentation

Experiment design in hours, not days

A modular A/B testing platform with live Bayesian readouts that show the team when a test has reached a decision.

bayesian readoutsself-serve designdays → hours

full story

Context

Every experiment was hand-crafted: sample sizes computed ad hoc, readouts assembled manually, and every decision waited on an analyst to declare the test done.

What I did

Architected a modular platform: a self-serve sample-size calculator at the design stage, automated pipelines feeding each running test, and live Bayesian readouts on top.

How

The Bayesian layer turns "is it significant yet?" into a visible decision signal — the dashboard itself shows when a test has accumulated enough evidence to call, taking the analyst out of the loop for standard readouts.

Signal

Experiment design dropped from days to hours, and readouts run self-serve. The modular core makes new test types a configuration task, not a rebuild.

probabilistic estimation

The reserves method that became the standard

A Monte Carlo methodology for probabilistic reserves estimation at an oil major — pricing geological uncertainty instead of betting on a single number. Still the company standard, four years after I left.

Monte Carlo · P10–P90$100M+ de-riskedcompany-wide standard

full story

Context

As the sole quant for the exploration division of a top-3 oil major's R&D center, I inherited a practice of point estimates — single confident numbers standing in for deeply uncertain geology.

What I did

Designed and operationalized a Monte Carlo–based probabilistic reserves methodology, and built the regression models — 90%+ accuracy, 16 points over legacy baselines — that underwrote drilling decisions.

How

Automated classification and clustering pipelines cut probabilistic evaluation cycle time by ~75%, making the method fast enough to become the default rather than a special-occasion exercise.

Signal

Adopted as the company-wide standard and still in production four years post-departure. The models backed $100M+ of capital across 18 exploration wells at a 100% drilling success rate; risk-adjusted frameworks prevented $3M+ in unprofitable license acquisitions.

background

Two domains, one method

2024 — Present

Head of Product Analytics · 01tech — built and run a 20+ person multidisciplinary analytics function (analysts, BI, ML, data engineering) across several consumer products; hired 12–15 analysts from 200+ interviews, promoted two seniors into team leads.

2023 — 2024

Senior Product Analyst · 01tech — on a flagship product, reframed customer care from a cost line into an ROI-run function; turned support noise into a structured, real-time product feedback signal.

2018 — 2022

Data Scientist · Gazpromneft STC — R&D arm of a top-3 oil major; sole quant for the exploration division. Models behind 18 wells at a 100% drilling success rate, and a Monte Carlo reserves methodology still in production as the company standard.

Off duty: travel and photography, tabletop games (I once built a board-game recommender just to settle arguments), and long walks. Based in Israel, working in English and Russian.