AI Engineer | LLM and RAG, Computer Vision, Audio | PyTorch | I measure what I build

I build end-to-end AI systems across language, vision and audio, and I measure them: a baseline, a fixed sample, and the failures read one by one.
My experience includes medical imaging research, applied deep learning, and building production-oriented AI applications.

A Claude Code skill that makes a Hugging Face model start fast on a free Kaggle GPU notebook.

• SAM-Audio large (6.9 GiB checkpoint, fp16) on a T4: build + load weights + copy to GPU from 96 s to 4 s, start to model on GPU from 158 s to 67 s, host RAM peak from 15.9 GiB to 8.5 GiB
• Each cause found with a probe on a free CPU notebook: random init of weights the checkpoint then overwrites, torch.load copying the whole file, a checkpoint read by one thread
• The fast path gives the same model: 1449 tensors compared, 0 differ

Repository: kaggle-fast
Technologies: PyTorch, Kaggle, uv, Claude Code

A document assistant tuned one change at a time on 150 questions over annual reports (FinanceBench), with every failure read.

• Writing the file name on every passage: right report in the top 5 from 70.7% to 84.7%, at no cost per question
• A bigger embedding model was worth 7 points alone, and 2 after two cheap fixes
• An LLM judge marked 93.85 wrong against 93.86: the grader gets checked too

Code is private for now; the findings go out as a LinkedIn series.
Technologies: Python, FastAPI, llama.cpp, hybrid search, rerankers

A page for a content creator with a chatbot that answers the audience as the creator, grounded in a profile they write themselves.

• Every statement in the character profile carries a cite to the creator's own words; without one the bot says it does not know
• Audience chat is read-only (one recall tool); edits from the admin console need an explicit confirm
• LLM, embeddings and text-to-speech each run locally or through a hosted API, chosen per role after a hardware probe

Repository: Self_representation_AI
Technologies: FastAPI, Next.js, llama.cpp, Qwen TTS, Docker Compose

Generative model for synthesizing CT from MRI to support radiotherapy dose planning.

• Developed preprocessing and paired dataset preparation pipeline
• Trained and evaluated Pix2Pix variants with and without a structural loss (NGF)
• NGF raised global SSIM from 0.7134 to 0.7315 (+2.5%) and Dice from 0.8826 to 0.8851

Repository / Report: mri-to-ct-synthesis-gan
Technologies: PyTorch, GANs, Medical Imaging

A local AI assistant supporting multiple LLMs and VLMs with document Q&A, voice input, and multimodal chat.

• Built a RAG pipeline with ChromaDB and multilingual embeddings for document retrieval over PDF, DOCX, XLSX, PPTX and more
• Integrated local speech-to-text transcription via faster-whisper with automatic language detection
• Designed a FastAPI backend with streaming inference, parallel model loading, and automatic context summarization
• Containerized with Docker and CUDA support for GPU inference

Repository: offline_multimodal_assistant
Technologies: Python, FastAPI, Streamlit, Transformers, ChromaDB, Docker

GitHub Trending to a Telegram digest to a LinkedIn draft, on free tiers only.

• Cloudflare Worker on a cron: top trending repos with an overview and two buttons, Draft Post or Skip
• Gemini reads the README and writes the draft; repos already seen are kept in KV and left out of later digests

Repository: trendhub
Technologies: TypeScript, Cloudflare Workers, Telegram Bot API, Gemini

Comparative study of segmentation architectures for lung region extraction.

• U-Net, DeepLabV3, and FCNN under identical training and evaluation settings
• Evaluation using Dice, IoU, and boundary metrics

Repository: U_Net-DeepLabV3-FCNN-lung-segmentation
Technologies: PyTorch, Computer Vision

Real-time pipeline combining YOLOv8 detection with SORT tracking.

• Extraction of spatial interaction descriptors (distance, direction, relative motion)
• Framework supports downstream tasks such as scene understanding or activity prediction

Repository: yolo-sort-spatial-relations
Technologies: YOLOv8, SORT, NumPy, OpenCV

End-to-end sentiment analysis workflow integrating language identification, translation, classification, and LLM-based summarization.

Repository: sentiment_analysis
Technologies: Python, Transformers, NLP

• Large Language Models, retrieval, and how to evaluate them
• Computer Vision and Image Understanding
• Medical Imaging and Scientific Computing
• Audio: speech, source separation
• Generative Modeling (GANs, diffusion models)
• Multimodal AI systems

Machine Learning:
Deep Learning, Computer Vision, NLP, Audio, Generative Models, LLMs, RAG and Evaluation, Representation Learning

Frameworks:
PyTorch, TensorFlow, Hugging Face, MONAI, Scikit-Learn, OpenCV, llama.cpp, ChromaDB

Systems & Tools:
Docker, Git, Linux, FastAPI, Streamlit, Next.js, Cloudflare Workers

Programming:
Python, TypeScript, C, C++