IT and Signals

ML/NLP/IR Engineer

Unit: Separate special e-support center
Category: Officer
Experience: No Experience — Training Provided

Responsibilities

training and bringing to production models in three related domains: ASR (Speech-to-text — fine-tuning modern architectures (CTC, RNN-T, TDT, attention-based) on noisy audio, PyTorch Lightning, NeMo, Hydra-config, multi-GPU), VAD (Voice activity detection — fine-tuning recurrent and CNN-based models on spectrograms, threshold optimization, ROC analysis), NLP / IR (text and search)
export of the model to ONNX, preparation of TensorRT-plans (FP16, optimization profiles, dynamic shapes), integration into NVIDIA Triton (ensemble, BLS, dynamic batching, instance groups)
construction of production services: FastAPI fronts, asynchronous workers on RabbitMQ / Kafka / S3 / Redis, integration into Flink pipelines
conducting thorough benchmarks: WER, F1 per-class, ROC-AUC, RTFx, throughput, A/B-tests with people markers (Prodigy)
reading papers, reproducing state-of-the-art approaches and adapting them to our data
data and models versioning (DVC + MinIO/S3), tracking of experiments (Weights & Biases), writing tests (pytest), passing code review

Requirements

5+ years of commercial experience in ML/NLP/Information Retrieval (CV experience is not required)
Python as the main language — confident knowledge of modern Python (3.11+), typing, async/await, writing clean supported code
PyTorch at the level of the author of custom training loops — not only Trainer.fit(), but also writing your loops, distributed training (DDP, gradient accumulation, mixed precision), understanding how loss functions under the hood work (CTC, RNN-T, contrasting losses)
at least one deep experience in one of the areas: ASR / sequence-to-sequence, classic NLP on transformers and classical methods, or Information Retrieval (BM25, dense retrieval, MinHash / LSH, ranking)
the ability to read and reproduce papers — take a model or approach from arXiv, understand trade-offs, adapt to your data, and not just run ready-made scripts
production engineering — experience in bringing a model from a laptop to a service: export (ONNX / TorchScript), optimization of inference, building a REST / gRPC API, writing tests (unit + integration)
HuggingFace ecosystem — confident work with transformers, datasets, tokenizers, accelerate, evaluate, hub; experience fine-tuning, custom training loops, custom tokenizers (BPE, SentencePiece, WordPiece).
classic ML — sklearn, XGBoost, threshold tuning for imbalanced multi-label tasks, Precision / Recall / F1 metrics (macro/micro/per-class), ROC-AUC
Git, pre-commit, linters, CI — take these things as the norm, not as an obstacle
knowledge of English at the level of reading technical documentation and papers without difficulty

What We Offer

Service under contract
small team, low bureaucracy, the decision is made by an engineer who knows the context
each change in model or infrastructure has clear benchmarks before/after — the unit does not "feel better" but measures
Production-mindset from the beginning: any model from the laptop should have a plan to access the server
availability of documentation, architecture diagrams, honest post-mortems, writing-it-down culture
transfer of active military personnel with the consent of the immediate commander
the possibility of service for candidates who, according to the conclusion of the VLC, are suitable for service in military support units, TCC and SP, higher educational institutions, training centers, etc.
office service in Kyiv, 5 days a week