ML/NLP Developer

ML/NLP Developer (Python + Go / Voice AI / LLM-dialogues) 🎙🧠

Vacancy ID: MLn_Fod Rate: to be discussed. Grade: Middle+ / Senior. Workload: Full-time. Format: Remote. Location/Citizenship: RF.

❗️ CRITICALLY IMPORTANT REQUIREMENT: Possession of a military ID or draft registration certificate is mandatory!

🏢 About the project: MTS. Development and evolution of a voice dialogue assistant. The key task is the phased transition of the system from rigid scenario graphs to a flexible, LLM-based dialogue with minimal latency.

🚀 Key Tasks: — Phased transition of the system from scenario graphs to LLM-based dialogue. — Debugging the voice pipeline: identifying causes of high latency, LLM hallucinations, issues with audio transmission, and graph errors. — Prompt architectural decisions and their independent implementation in code. — Support and debugging of few-shot scenario classification (Qwen, BERT), correction of erroneous intent triggering. — Finetuning local LLMs as needed (entity extraction, instruction adherence). — Integration and support of ASR (GigaAMv2), retraining and configuring inference for TTS and voice cloning systems.

🎯 Mandatory Requirements (Hard Skills):

NLP and LLM (CRITICAL!): — Solid experience with LLM/SLM (Few-shot learning, prompt engineering) and BERT (scenario and intent classification). — Practical experience in finetuning LLMs (PEFT, LoRA/QLoRA) for specific application tasks. — Solid LLM debugging skills: ability to quickly understand the cause of an incorrect response and fix it via prompt, hyperparameter tuning, or finetuning.

Voice Stack (CRITICAL!): — Experience in retraining and optimizing TTS models. — Experience with and integration of modern ASR systems (specifically, GigaAMv2), understanding of VAD operation. — Practical experience with real-time audio streaming (streaming ASR/TTS).

Python and Inference Optimization: — Proficient in Python (PyTorch, HuggingFace, inference frameworks). — Experience in optimizing inference for low-latency (vLLM, TensorRT-LLM, Triton Inference Server, ONNX Runtime).

Go (CRITICAL!): — Understanding of service design, SOLID, DRY, and KISS patterns. — Type system in Go (strings, numbers, slices, maps). — Synchronization primitives and concurrency (goroutine synchronization, working with channels). — Pointers and working with them. — Understanding of Observability principles.

💡 Would be a significant advantage: — Experience with gRPC contracts. — Ability to configure infrastructure, knowledge of k8s / inventory. — Deep understanding of the network stack (TCP).