Evolutionary algorithms
LLM-guided evolution over program spaces, where a population of candidate programs is mutated, scored and selected across generations.
AI Researcher at Fusionbrain Lab
Iana Kulichenko
I work on multi-agent LLM systems, evolutionary search for alpha signals on financial markets, generation of 2D and 3D apartment layouts, and medical image segmentation. I design the architecture and train the models these systems run on.
PhD candidate in Computational and Data Science and Engineering at an AI research institute. Co-founder of Fiber Pipe.
- genetic programming
- MAP-Elites
- walk-forward validation
- deflated Sharpe ratio
- multi-agent orchestration
- agent prompting
- tool use
- graph RAG
- mid-training
- SFT
- LoRA and QLoRA
- DPO
- GRPO
- constraint satisfaction
- GAN inversion
- Textual Inversion
- MedSAM
- scene graphs
- genetic programming
- MAP-Elites
- walk-forward validation
- deflated Sharpe ratio
- multi-agent orchestration
- agent prompting
- tool use
- graph RAG
- mid-training
- SFT
- LoRA and QLoRA
- DPO
- GRPO
- constraint satisfaction
- GAN inversion
- Textual Inversion
- MedSAM
- scene graphs
- Python
- C++
- PyTorch
- TensorFlow
- Hugging Face
- Docker
- Linux
- Weights & Biases
- MLflow
- SQL
- R
- Git
- LaTeX
- Python
- C++
- PyTorch
- TensorFlow
- Hugging Face
- Docker
- Linux
- Weights & Biases
- MLflow
- SQL
- R
- Git
- LaTeX
Focus
Multi-agent LLM systems
Multi-agent systems for applied tasks: role decomposition, orchestration and tool use, graph RAG, and training of the agent models themselves. Mid-training on domain corpora, post-training with SFT, LoRA and QLoRA, DPO and GRPO.
Generative models
Diffusion models and GANs for image synthesis and editing, including latent-space inversion and subject personalization.
Multimodal LLMs
Adapting vision-language models to new tasks with parameter-efficient fine-tuning, supervised fine-tuning and reinforcement learning post-training.
Projects
Evolutionary alpha search on markets
An LLM proposes and mutates factor expressions for Russian equities, generation after generation. Every candidate is scored by walk-forward backtests over a decade of daily data and screened against multiple testing with the deflated Sharpe ratio, probability of backtest overfitting and the model confidence set, so that survivors are not artefacts of the search itself.
Multi-agent floor plan generation
A multi-agent LLM system recommends furniture selections and placements within defined room boundaries, then renders realistic 3D rooms.
StyleGAN-2 encoder optimization
Encoder optimization with an analysis of image inversion and editing methods, raising the quality of generated and edited images.
Multimodal LLM fine-tuning
Parameter-efficient fine-tuning, including QLoRA, applied to InternVL2-1B to lift performance on the temporal tasks of MVBench.
SegMed
Brain tumor segmentation on the LGG MRI dataset. MedSAM measured against U-Net, U-Net++, PAN and DeepLabV3+ baselines, reaching mean IoU 0.644, then extended with FeatUp for feature upsampling and GAFL for adaptive frequency filtering.
Personalized image generation
Embeddings tuned through Textual Inversion in Stable Diffusion, creating custom tokens for personalized subjects.
Panorama stitching
SIFT and ORB feature matching with camera calibration, producing high-precision panoramic images.
Publications
IEEE Sensors Letters, 2025
Machine Learning-Driven Compensation for Non-Ideal Channels in AWG-Based FBG Interrogator
An exponential-regression calibration of a photonic integrated AWG interrogator reaches 3.17 pm RMSE against 7.11 pm for the segmented analytical model, and holds accuracy below 5 pm across an extended 2.9 nm span without refitting.
Neurocomputers: Development, Application, No. 5, 2023
Generative image models for augmenting the training data of a face detector
Synthetic faces produced by generative models are added to the training set of a face detector to cover poses and lighting the collected data misses.
Применение генеративных моделей изображений для аугментирования данных обучения детектора лиц. Н. А. Андриянов, Я. В. Куличенко. Журнал «Нейрокомпьютеры: разработка, применение», издательство «Радиотехника».
radiotec.ru
Listed by the Higher Attestation Commission
A. S. Popov Society, 2023
A study of metric algorithms for face recognition
Comparison of metric learning approaches on the face recognition task, measured by identification accuracy across embedding distances.
Исследование метрических алгоритмов в задаче распознавания лиц. Куличенко Я. В., Уткин Д. С., Фан А. Ч., Матусков Н. И., Лопаткин И. М., Андриянов Н. А.
A. S. Popov Society, 2023
Faster face detection and identification through a motion detector
A motion detector gates which frames reach the recognition stage, cutting the work the face pipeline has to do on static video.
Повышение скорости детекции и идентификации лиц на основе детектора движения. Уткин Д. С., Куличенко Я. В., Андриянов Н. А.
Patents and IP
Three items registered with Rospatent, the Russian federal intellectual property service, filed under Куличенко Яна Владимировна.
RU 2861310 C1
Method for interrogating fiber Bragg gratings through an arrayed waveguide grating demultiplexer on a photonic integrated circuit
Machine learning recovers the reflected Bragg wavelength from per-channel optical power, replacing the analytical fit that a non-ideal channel response breaks. Registered 4 May 2026.
RU 2026613062
Data processing software for fiber Bragg grating sensors
Converts the analog photodiode signal to digital, averages readings over a time window and computes measurement error including the standard deviation of the incoming voltage. Written in C++. Registered 3 February 2026.
RU 2025681052
Interrogator user software
Real-time acquisition, processing and display of fiber Bragg grating sensor data, with adaptive analysis and storage of historical values. Written in TypeScript. Registered 11 August 2025.
Talks
Generative AI agents
Gitex Global 2025, Dubai World Trade Centre. gitex.com
Machine Learning-Driven Compensation for Non-Ideal Channels in AWG-Based FBG Interrogator
IEEE SENSORS 2025, Vancouver, Canada. Presented online. ieee-sensorsconference.org
State of the Art YOLO Model in Object Recognition Tasks
XIV International Scientific Student Congress, Moscow. Second place in the competition.
Modern Text-to-Image Generation Technologies
International Scientific and Practical Conference of Students and Postgraduates, Moscow.
Background
Experience
- 2025 to now Research Engineer, Fusionbrain Lab LLM-based generation of 2D floor plans. RL and supervised fine-tuning for LLM post-training.
- 2024 to 2025 Data Scientist, Fusionbrain Lab StyleGAN-2 encoder optimization, evaluation of image inversion and editing methods.
- 2023 to now Co-founder, Fiber Pipe Pipeline monitoring in the Arctic through computer vision and fiber-optic sensing. First place at the EnergyTechnoHub incubator in Saint Petersburg, winner of the Triple Point pitch competitions.
Education
- 2025 to 2028 PhD, Computational and Data Science and Engineering AI research institute. Generative AI systems for full-apartment layout and interior design.
- 2023 to 2025 MSc, Data Science Segmentation, multimodal models, multi-agent systems, generative models.
- 2019 to 2023 BSc, Applied Mathematics and Informatics Mathematics, statistics, classical and deep machine learning, big data.
Open to research collaboration
Generative AI, multi-agent systems, evolutionary algorithms, alpha signal search on financial markets.