ML engineer for social discovery products

Dating Group; United States; À distance; USD 140 000 - 210 000; Plein temps

Avantages offerts par ce poste

Vacation 28 calendar days

Wellness days 7

Referral bonuses up to $5000

Training reimbursement 50%

English lessons discount

Health benefits up to $1000

Workspace provided

Co-working reimbursement every 3 years

Résumé du poste

Social Discovery Group ищет ML-специалиста, который задаст направление разработки, возглавит адаптацию моделей и обеспечит интеграцию LLM в продукты.

Требуется глубокое знание Python, Transformers и практический опыт обучения и тонкой настройки моделей; знание SFT/LoRA/PEFT приветствуется. Условия удалённой занятости, CET-время, бонусы и компенсации.

Qualifications

• Глубокие знания Python и опыт работы с трансформерами.
• Опыт обучения и точной подстройки LLM (SFT, LoRA/PEFT).
• Умение строить наборы данных: очистка, фильтрация, дедупликация.
• Опыт оценки LLM: офлайн и парные оценки, субботние тесты.
• Проектирование​​ агентных
• инструментов вызова функций, маршрутизации и памяти.
• Интеграция LLM в продакшн: API, стриминг, батчи, траты на latency.
• Хорошее владение русским языком; готовность к CET.

Responsabilités

• Set the ML direction: выбор стратегических ставок с привязкой к бизнес-целям.
• Руководство пост-обучением LLM и адаптацией моделей от набора данных до финальной настройки.
• Дизайн и сборка систем на базе агент-гарнитов и инструментов.
• Разработка рабочих процессов multi-agent и оценочных фреймворков.
• Улучшение качества диалогов, памяти, персоны и безопасности.
• Развертывание производственных LLM-систем с учётом latency и стоимости.

Connaissances

Python expert

Transformers

Training и fine-tuning LLMs

SFT

LoRA/QLoRA/PEFT

Preference optimization

Dataset construction

LLM evaluation

Agent harness design

Production LLM integration

Observability for agents

Transformer architecture knowledge

fp32/fp16/bf16, CUDA memory

Git, Docker, Linux, MLOps

Fluent Russian

CET hours readiness

Experience beyond text (multimodal)

Outils

Git

Docker

Linux

MLOps basics

Description du poste

Описание:

Social Discovery Group (SDG) is a group of social discovery companies. SDG solves the problems of loneliness, isolation, and disconnection - transforming virtual intimacy into the new normal. SDG’s products redefine the way people interact and connect with one another. Our portfolio includes social entertainment platforms designed to connect people online across different cultures and regions of the world.

Задачи:

Set the ML direction: choose which strategic bets to make and tie each one to clear business needs. Drive LLM post-training and model adaptation end to end, from dataset construction to fine-tuning and preference optimization. Design and build agent harnesses and tool-using LLM systems. Shape multi-agent development workflows that make the team faster and more effective. Build evaluation frameworks and LLM-as-a-judge systems that reliably predict real-world performance. Improve dialogue quality across memory, emotional intelligence, persona consistency, safety and anti-loop behavior. Ship production systems that directly shape user experience, retention and monetization.

Требования:

• Python expert level: clean code, typing, async, multiprocessing, profiling, debugging, notebooks + production code.
• Transformers.
• Training and fine-tuning LLMs.
• SFT, LoRA/QLoRA, PEFT.
• Preference optimization: DPO / ORPO / SimPO / CPO / GRPO or willingness to quickly go deep.
• Building datasets for LLM training: cleaning, filtering, deduplication, preference pairs, synthetic data, hard negative mining.
• LLM evaluation: offline evals, pairwise evals, LLM judges, human annotation.
• Agent harness design: tool calling, function calling, structured outputs, retries, routing, state management, memory/context, guardrails.
• Production LLM integration: APIs, streaming, batching, fallbacks, cost/latency trade-offs.
• Observability for agents: traces, tool‑call logs, transcript analysis, failure clustering.
• Strong understanding of transformer architecture and modern LLM behavior.
• Accuracy with dtype/device/batching: fp32/fp16/bf16, quantization, CUDA memory, avoiding dtype mismatch bugs.
• Git, Docker, Linux, basic MLOps.
• Fluent Russian, ready to work in CET (±2) hours.
• Будет плюсом: experience beyond text (computer vision, image generation, multimodal); long-running conversations and character consistency; vLLM / TGI / SGLang; DeepSpeed / FSDP / Accelerate; quantization; safety classifiers

Условия:

REMOTE OPPORTUNITY to work full-time The initial pay level or pay range for this role will be shared with candidates during the recruitment process and before the commencement of employment Vacation 28 calendar days per year 7 wellness days per year (time off) that can be used to deal with household issues, to lie down and recover without taking sick leave Bonuses up to $5000 for recommending successful applicants for positions in the company 50% payment for professional training, international conferences, and meetings Corporate discount for English lessons Health benefits. According to the paychecks, if you are not eligible for corporate medical insurance, the company will compensate you with up to $ 1,000 gross per year per employee. This can be spent on self-purchase of health insurance or on doctor’s fees for yourself and close relatives (spouse, children) Workplace organization. The company provides all employees with an equipped workplace and all the necessary equipment (table, armchair, wifi, etc.) in our offices or co-working locations. In the other locations, the company provides reimbursement of workplace costs up to $ 1000 gross once every 3 years, according to the paychecks. This money can be spent on the rent of the co-working room, on equipping the working place at home (desk, chair, Internet, etc.) during those 3 years Internal gamified gratitude system: receive bonuses from colleagues and exchange them for our merchandise, team building activities, massage certificates, etc.