КазахстанData Science, Product ML Engineer (Personalization & Monetization)
1 мес. назад
КазахстанOnsite
machine learningpythondata analysisrecommendation systems
Build and own production machine learning models for personalization and monetization to drive user-specific product decisions.
О компании
- Higgsfield AI is the fastest-scaling generative AI company in history, hitting $500M in annual revenue run rate, 25M+ 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.
Условия
- Competitive base salary in USD , based on your experience, skills, and the scope of the role.
- Equity participation through the company’s stock option program, giving you the opportunity to share in Higgsfield’s long-term growth.
- Relocation support to Almaty for candidates moving from another city or country.
- A highly collaborative, fast-paced environment where you can work directly with experienced leaders and have a meaningful impact on the product and company.
- Opportunities for professional growth, ownership, and career development as the company scales.
- Company-provided equipment, meals, transportation, or other office benefits.
- This is a fully on-site role based in our Almaty office . Our team works from the office five days per week for the full working day . We believe in-person collaboration is an important part of how we move quickly, solve complex problems, and build strong teams.
Другое
- Right now every user sees roughly the same Higgsfield. The same effects in the same order, the same paywall, the same offer, the same lifecycle emails - across 25M+ people who want completely different things from us.
- This role makes the product decide per user.
- You own the models behind that: what we recommend, who gets which offer, who we spend retention effort on, and what each of those is worth. Not a research agenda - production models on live traffic, measured in revenue, retention and margin.
- You make sure:
- What a user sees next is ranked for that user, not for the average of everyone
- Discounts and offers go to the people whose behaviour they actually change, and nowhere else
- We know who is about to leave early enough to do something about it and we know whether the something worked
- Every model in production has a holdout behind it, so we can always state what it is worth
- Build the ranking systems for our content surfaces: effects and presets, templates, models, prompts, and what we show a user next after a generation completes.
- Solve cold start properly - a brand-new user with one onboarding signal and no history is our single most common case, not an edge case.
- Handle the things that make recsys hard in the real world: position bias, popularity feedback loops, exploration vs exploitation, and the fact that a ranker trained on logged behaviour learns to reproduce the ranking it was trained on.
- Rank against the objective we actually want - a completed, kept, shared generation - not the click.
- Own incrementality on offers: discounts, credit vouchers, trials, plan upgrade prompts, win-back campaigns.
- Model uplift, not propensity. Targeting the users most likely to convert spends margin on people who would have converted anyway; the entire value of this work is finding the users whose decision the offer changes.
- Design the randomized holdouts that make uplift trainable and measurable in the first place, and keep them running permanently.
- Respect the guardrails: an offer model optimizing conversion alone will happily find the accounts that convert at negative gross margin, and it will find abuse rings first. Margin and abuse signals are constraints on the objective, not a later cleanup.
- Work with Legal on what may be personalized. Personalized pricing and discounting touches consumer-protection and consent rules in several of our markets - you should want that conversation, not route around it.
- Build churn and downgrade prediction for subscribers, and repeat-purchase propensity for credit-pack buyers.
- Know the difference between a model that predicts churn and a model that reduces it. A high AUC earned by detecting users who already stopped using the product is worth nothing.
- Watch for leakage relentlessly - cancellation-adjacent features will hand you a beautiful offline number and a useless system.
- Pair every propensity model with an intervention and an experiment. The deliverable is a retained user, not a score.
- Turn what users actually make into features the models above can use: use-case and intent labels over prompts, input assets, output assets, model and parameters.
- Content understanding here is instrumental — a taxonomy exists so ranking knows what a preset is for and so an offer knows which use case a user is stuck on. Multimodal, because a large share of our generations carry almost no prompt text and the intent lives in the uploaded image.
- Mine failed, abandoned and refunded generations: they are the strongest churn and unmet-demand signal we have, and success-only training data has survivorship bias built in.
- Own models end to end: problem framing, features, training, offline evaluation, serving, monitoring, retraining.
- Ship behind experiments. Offline metrics decide what to try; online experiments decide what stays.
- Monitor for drift and degradation - new model launches, pricing changes and seasonality all move the ground under a deployed model.
- Write down what each model is worth, in money, and keep that number current.
- Experience shipping ML that served live user traffic and changed a business metric. Notebooks and offline benchmarks are not this.
- Depth in at least two of: recommender systems / learning-to-rank, uplift & causal ML, churn or propensity modelling, real-time personalization.
- Genuine causal literacy: randomized holdouts, incrementality, Qini/uplift evaluation, selection effects, why a lift measured before-and-after is usually not a lift.
- Strong Python and SQL; comfortable with gradient boosting and neural ranking, and with the boring parts - feature pipelines, training/serving skew, latency budgets, retraining cadence.
- Product judgment: you can name the decision and the metric before you pick the model, and you know when the answer is a rule rather than a model.
- Comfort with images and video as data, or clear appetite to get there fast.
- Pragmatism. A heuristic shipped this month that lifts conversion beats a two-quarter platform. Then you replace the heuristic.
- Clear written and spoken English, B2+.
- Recsys / ranking engineers from consumer products at scale
- Growth or monetization ML from subscription, gaming, fintech or e-commerce - anywhere offers and churn are modelled with real money attached
- Causal inference / uplift specialists who ship systems rather than studies
- Applied scientists who own their models in production
- This role is not a fit if you:
- Want to train or fine-tune generative video models - that's our R&D ML Engineer roles, and they're open
- Optimize offline metrics and hand the model to someone else to deploy
- Would target an offer at the users most likely to buy and call it personalization
- Report a lift from a before-and-after comparison
- Need clean labelled data and a feature store to exist before you can start
- Want predictable 9–5 workdays