КазахстанData Scientist, Antifraud
2 нед. назад
КазахстанOnsite
machine learningdata analysisstatistical modeling
Build and operate antifraud systems to detect and prevent abusive behaviors in subscription-based generative video services.
О компании
- is the fastest-scaling generative AI company in history, hitting $1B in annual revenue run rate, 30M+ users worldwide, 6M+ generations per day, and powering 390 of Fortune 500 brands.
- We're building at the absolute frontier of AI-powered video creation and next-generation creative tools. Joining Higgsfield means becoming part of a high-impact team shaping the future of AI-native experiences, at a company that isn't just moving fast, but rewriting what fast looks like.
Как откликнуться
- We move fast:
- Screening call (30 min)
- Hiring-manager screen (30 min)
- Business case (1 hour)
- Practical home task (7 days)
- Team interview (60 min)
- Paid on-site trial (1 month)
Другое
- Generative video has a problem most subscription businesses don't: every unit we sell costs us real compute. A stolen password costs a streaming service nothing. An account farm on an unlimited generation plan costs us GPU-seconds, at scale, every hour it runs.
- We have found organised operators running hundreds of accounts on a single card, reselling unlimited generations, and harvesting per-account promotional grants at industrial scale. We have also found that most of our negative-margin accounts are not abuse at all — they are ordinary heavy users on an unlimited plan, which is a pricing problem wearing a fraud costume. Telling those two populations apart, at account level, before the money is gone, is this job.
- You own the question of who is taking value they did not pay for, and the system that stops them.
- You make sure:
- We catch abuse before the compute is burned, not in a post-mortem that reconciles the loss
- Every enforcement action has a measured precision, a measured cost of being wrong, and a way back for the customer we got wrong
- The difference between fraud and unprofitable pricing is a number we can defend, not a vibe
- When operators adapt — and they will, within days — we find out from a monitor, not from the monthly margin review
- Own the account-abuse scoring system end to end: signals, weights, thresholds, the scoring job, the monitoring, and the recalibration when it decays.
- Build detection that fires early . A rule that catches 88% of the money on day 11 is worth far less than one that catches 60% at checkout, and you should be able to say why in dollars.
- Work the two halves that actually matter together: linkage (who is connected to whom) and economics (what are they consuming). Either one alone produces a system nobody can switch on — linkage without economics flags families sharing a card, economics without linkage flags our best customers.
- Treat cost as a gate, not a label. On an unlimited plan, being gross-margin negative is normal.
- Find rings, not just accounts: shared payment instruments, IP and ASN concentration, email-stem families, synchronised registration bursts, behavioural fingerprints, automation signatures.
- Do it without a reliable device fingerprint — we do not have one today, and part of this job is telling us what it would be worth and what it would cost.
- Know where graph methods break. Transitive closure through a shared card will happily merge thousands of unrelated people into one "ring"; we have done exactly that and thrown the result away.
- Design graduated, reversible actions — throttle, rate-limit, step-up verification, hold, pre-grant refusal, manual review queue, ban — and match the severity to the confidence.
- Own the false-positive budget explicitly: how many paying customers are we willing to inconvenience to save a dollar of compute, and what is the appeal path for the ones we get wrong.
- Build the pre-grant gates. The cheapest abuse to stop is the kind we decline to enable in the first place — a trial we do not grant costs nothing to claw back.
- Partner with Payments on Stripe Radar, chargebacks and dispute economics; with Support on the queue your model creates; with Legal on what our Terms actually entitle us to do.
- Quantify what the system is worth in money — COGS avoided, net of the revenue you destroyed and the support cost you created. Keep that number current.
- Run holdouts even on enforcement, wherever it is legal and safe to do so, because otherwise you will never know what your rules are worth.
- Monitor for adaptation , not just drift. Model decay here is caused by an adversary, and it looks different: a signal stops firing, a ring changes shape, a cohort's behaviour shifts inside a week.
- Watch your own inputs. A detection system that silently loses a feature degrades to a system that cannot reach its own threshold — and reports nothing wrong while doing it.
- Ship scores and reasons, never scores alone. An analyst, a support agent or an auditor must be able to read why an account was actioned.
- Produce evidence that survives outside the company. Our abuse work has already been used in an external audit and in Terms-of-Use enforcement; write for that standard.
- Work in SQL, Python and BigQuery on raw event, payment and generation data. We use AI heavily — resourcefulness beats syntax.
- Experience building abuse, fraud or trust-and-safety detection that actually took action on real users — payments fraud, promo abuse, multi-accounting, bonus abuse, account takeover, marketplace or gaming abuse. Anti-money-laundering and credit-risk backgrounds transfer if the modelling was yours.
- Genuine understanding of the adversarial setting: that your counterparty responds, that your training labels are generated by your own past decisions, and that offline AUC on a censored label set is close to meaningless.
- Strong SQL and Python on raw event data without help: joins across payments, usage and identity, window functions, cohort replay. Gradient boosting, and the judgment to know when a two-condition rule beats a model.
- Comfort with graph and linkage thinking — connected components, shared-attribute clustering, and an instinct for when a link is evidence and when it is coincidence.
- Precision-first evaluation instincts: precision at the operating point, cost-weighted error, and why you would take 20% recall at 90% precision over the reverse in an enforcement system.
- Unit-economics literacy: gross margin per account, what compute actually costs us, why a negative-margin customer may be entirely legitimate.
- The judgment to escalate rather than act when the evidence is thin, and to say so in writing.
- Clear written and spoken English, B2+.
- Fraud / risk data scientists from fintech, payments, marketplaces, gaming or crypto
- Trust & safety and platform-integrity teams who owned detection and enforcement policy
- Growth or product data scientists who ended up owning promo-abuse because nobody else would
- Analysts from payment processors, PSPs or chargeback/dispute operations who moved into modelling
- This role is not a fit if you:
- Want to build content-moderation or NSFW classification — that is a different function here
- Would ship a ban list without knowing its precision
- Optimize AUC on a label set your own enforcement created, and call it validated
- Need a labelled fraud dataset to exist before you can start
- Treat every negative-margin account as an abuser
- Are uncomfortable that being right 90% of the time means being wrong about real, paying, legitimate people — and that you own what happens to them
- Want predictable 9–5 workdays
- Competitive salary in USD
- Equity: participation in the company's stock option program
- On-site role in Almaty, Kazakhstan
- Relocation support (flight + temporary housing)
- Flat structure, high autonomy, fast career growth