I build analytics that runs itself and products that ship.

An analytics team with 20 AI-agent skills

Input

Routine requests: finding the right table, standard analyses, editing reports for the business.

Model

Claude Code with skills, agents and machine-readable cards of every table.

Guards

A hook that blocks destructive SQL and a pre-commit scanner for secrets.

Impact

Handed over to the team's shared repository in September 2026.

Dashboards built by an agent

Model

The skill writes Tableau workbooks without Tableau Desktop, publishes them, checks the render and the CSV, and rolls back from a backup on failure.

Impact

Production dashboards are built with it, including the VIP bonus limits board; shared with other analytics teams.

Grew two analytics team leads

How

Hiring, development and a shared base of tasks with solution code: new analysts start from working examples, not from scratch.

Churn model the call centre acts on every morning

Input

42 features: sessions, 7- and 30-day trends, negative experiences, payments.

Model

CatBoost: ROC-AUC 0.756 vs 0.675 for the ML team's model on the same 29.6k players.

Action

Ranks who will leave in the next 30 days, so the call centre calls them first.

Guards

Training gate and drift monitor added after a degradation episode; calibration within ±4 pp; carried over to a second business line.

Impact

Precision on the top 20k: 55.7% vs 46.3%; ROC-AUC 0.85 in production.

Finding the future VIPs the main model misses

Input

Payment behaviour on day 14 and day 21 after the first payment, only for players the ML model did not flag.

Model

CatBoost, ROC-AUC 0.87–0.89. Players in the top 1% of the forecast become VIPs 16 times more often than average.

Action

VIP managers get 27% more leads every day, none of them overlapping the main model.

Guards

Time-based split; the first production run passed 295 of 295 selection checks.

Impact

The day-21 checkpoint added another +40% leads at the same precision.

Personal contact algorithm for VIP players

Input

Each player's personal norm of breaks between deposits and play.

Model

A per-player threshold instead of a fixed 14-day schedule.

Action

When a player is silent longer than their norm, their VIP manager gets a task the same day.

Guards

A/B with ANCOVA, SPRT and intention-to-treat; a second test across three regions runs since October 2026.

Impact

+6% deposit days in the A/B test (p < 0.01); +11 pp return within 3 days when the manager reaches out on the signal day; in another region the same result with 12% fewer contacts.

Predicted Value Score instead of a tier that only grows

Input

19 features on payments, withdrawals, play and recency over 30–365 days, across the whole player base.

Model

Gradient boosting predicts 90-day value and maps it to a V0–V8 scale.

Action

Re-segments every player monthly; managers and bonus limits run on the new scale.

Guards

Out-of-time backtest on 4 cuts, a quality gate before each upload, monthly fact-vs-forecast alert.

Impact

At equal list size it captures 48–49% of quarterly revenue vs 38–39%. Handed to the ML department.

Bonus limits built on real game margins

Input

Actual margin by 9 game categories, net of fees and partner share, adjusted for bonus wagering.

Model

Theoretical net revenue per player instead of book margins — some games were off by a factor of three.

Action

Daily bonus limits for every player, shown to managers in an internal tool and a dashboard.

Impact

$7–11M a year in estimated effect: every player's bonus limit is now based on what that player actually brings in.

Answering-machine detection on live calls

Input

The first 3 seconds of every outbound call.

Model

CatBoost on 184 acoustic features, served by FastAPI and wired into the telephony stack.

Action

Drops the call when a machine answers, so operators talk only to people: ~30 s saved per such call, and machines take 20–40% of calls.

Guards

Fail-safe: any error counts as a human. Retraining removed the 10 s vs 3 s gap between training and production audio.

Impact

Recall 79% → 93%, F1 85% → 95%, false positives 3.6% → 2.9%. Handed over to the VoIP team.

AI checks support chats: more than half with no human

Input

Support chats that go to quality control: was the issue solved, empathy, critical errors, grammar.

Model

An LLM checks the chat against the QA criteria, points the manager to the operator's mistakes and closes flawless chats on its own.

Action

QA managers check the rest with a ready AI hint: where the operator went wrong and against which criterion. Disagreements feed the next prompt version.

Guards

Personal data is masked before a chat reaches the model. If the AI cannot check a chat, a person does, and the pipeline keeps running.

Impact

More than half of support chats are checked by AI with no human; QA managers check the rest with its hints.

Alerts that read support chats

Input

Support chat texts, payment complaints per hour, chat availability, call-centre metrics.

Model

Robust z-scores against the same weekday, a burst rule and detection of new error codes.

Action

Product breakages and payment spikes go to Telegram every hour, without anyone opening a dashboard.

Measuring the conversation, not the campaign

Model

Everyone is called; randomisation happens among those who picked up, and in control the call is dropped. The test measures the effect of the conversation itself (LATE).

Impact

A 3.4 pp effect is visible in 7 days, because the test runs only on people the call can actually influence.

Experiments that stop themselves

Model

Sequential probability ratio test: the experiment ends as soon as the evidence is enough, not on a fixed date.

Impact

Introduced across the department; average test duration halved.

DataGlass

Market intelligence for iGaming

What it does

Traffic, demand signals, licences, payment methods and bonus offers for hundreds of operators, refreshed every month. SaaS with a 14-day trial.

PaysOrNot

Does the casino pay?

What it does

Accountability index and payout reliability built from 47,000+ player reviews across five platforms, with a country-level payments signal. Operators can answer complaints.

dalat.place

City guide to Dalat, Vietnam

What it does

Catalogue of places with search that understands synonyms and transliteration, live weather and its own analytics dashboard: returns, sources, funnel.

Mirror

AI bot with memory of the person

What it does

Telegram bot that catches self-deception. Its memory is a typed model of the person — values, fears, contradictions with quote and date, a map of patterns — written by a cheap LLM and read by a strong one. Memory decides the tactics: when to push, when to support, when to reach out first.

Nod

Control the computer with your face

What it does

The cursor follows your head, a wink clicks. 478 face landmarks, One Euro filter against tremor, calibration with a double blink, packaged as a macOS app.

Bienveille

Eldercare without cameras

What it does

A sensor on the electricity meter learns a person's daily rhythm and raises an alert when the morning has not started. Personal model checked on real smart-meter data from 13 homes of retirees.