Position: Senior NLP Engineer Company: Viktori-Tekhnolodzhy City: Astana Employment: hybrid Salary: from 1,000,000 KZT.

Vacancy Description: Responsibilities:

• Development and improvement of LLM pipeline Stage 1: cleanup of transcripts after diarization (Sortformer + Whisper Large V3), automatic speaker role identification and flagging of significant events in the dialogue;
• Development and improvement of LLM pipeline Stage 2: automatic generation of structured documents in DOCX/XML formats for a backend editing system and multi-stage compliance check according to domain rules (~30+ regulatory requirements);
• Fine-tuning and prompt engineering of LLMs (gpt-class and open-weight models) for domain-specific tasks in Kazakh and Russian languages; reducing hallucination rate and stabilizing outputs between runs;
• NER in domain-specific texts: dialogue participants, actions, temporal and spatial markers, regulatory references;
• Semantic segmentation of transcripts: splitting into meaningful episodes and highlighting key moments (formal procedures, petitions, pauses, question-answer);
• Development of evaluation methodologies and regression tests for LLM outputs; maintaining model benchmarks (accuracy according to domain rules, false-positive / false-negative, stability between runs);
• Creation of an LLM agent that interprets behavioral and emotional signals from CV and Audio modules and generates a coherent analytical text for the end-user;
• Working with legal ontologies and regulatory sources: Criminal Procedure Code of the Republic of Kazakhstan (including Articles 68, 77, 115), Criminal Code of the Republic of Kazakhstan, Constitution of the Republic of Kazakhstan; maintaining the knowledge base up-to-date with legislative changes;
• Retraining and optimization of ASR models (Whisper Large V3, Sortformer V4 for diarization) in Kazakh and Russian languages with adaptation to the acoustics of the target room, background noise, regional accents, and domain-specific vocabulary; ONNX/TensorRT export and deployment on Jetson;
• Development of a pipeline for extracting audio-prosodic features (F0/pitch, jitter, shimmer, intensity, speech rate, pauses, voice quality, eGeMAPS) and modeling psycho-emotional markers (stress, uncertainty, emotional tension) based on them;
• Development of the final multimodal fusion layer: integration of text embeddings (NLP), audio-prosodic features, and visual signals from the CV module (facial expressions, micro-movements, posture) into a unified interpretation of the speaker's psycho-emotional state;
• Generation of final analytical conclusions for the end-user in the form of structured multimodal annotations: dialogue fragment with prosodic context and behavioral markers, ranking of moments by significance level with indication of model confidence level.

Requirements:

• 3+ years in NLP/LLM; experience with production projects with end-to-end LLM pipelines;

• Deep knowledge of transformer and LLM architectures; experience in fine-tuning (LoRA / QLoRA / SFT) and instruction tuning for domain-specific tasks;

• Confident proficiency in HuggingFace Transformers, vLLM, llama.cpp; experience in inference optimization (quantization, batching, KV-cache); working with locally deployed models;

• Deep skills in building RAG: chunking strategies, vector databases (pgvector / FAISS / Pinecone), reranking, multilingual embeddings (including Kazakh);

• Experience in minimizing hallucinations: structured outputs, guardrails, constrained decoding; LLM evaluation (ragas, LangSmith, and similar);

• Experience in industrial development in Python (FastAPI / Flask, Docker, Git); understanding of the legal domain and working with KZ/RU languages;

• Experience in retraining ASR models (Whisper / wav2vec2 / Sortformer) on domain-specific data; working with speaker diarization, speaker embeddings, and noise reduction;

• Practical experience in extracting and modeling audio-prosodic and paralinguistic features (librosa, openSMILE, parselmouth / Praat, eGeMAPS); understanding of acoustic correlates of emotions and stress;

• Experience in multimodal fusion (early / late / hybrid fusion, cross-modal attention); working with Audio Transformers (AST, HuBERT, wav2vec2) and multimodal LLMs.

• Projects in LegalTech and digitalization;

• Real tasks, quick solutions, and space for initiative;

• Employment as staff, probationary period of 3 months;

• Schedule: 5/2, 9:00–18:00.

Contacts: Telegram recruiter contact