Javohir Hoshimov
Lead AI Engineer · Tashkent
I build AI that reaches real users. At SinoAI I work on a medical consultation assistant: a RAG pipeline over medical knowledge, real-time voice consultations over WebSocket, and the backend that runs them on Telegram, web and mobile.
Outside work I build things that see and measure: player tracking for football, heartbeat classification from a smartwatch, tumour segmentation in ultrasound.
Projects
-
FootCoach · 2026
Football analytics from one fixed camera. A GPU worker detects and tracks every player (RF-DETR + ByteTrack), maps them onto the pitch, and turns a match into distance, sprints, formations and space control. An AI analyst answers a coach's questions over the club's own data.
-
xatolab · 2026
IELTS practice that takes every wrong answer apart. 6,071 questions across 157 Listening and Reading tests: it grades, logs each mistake, finds the weakest task type and plans the next 20 days around it.
-
HealthConnect · 2026
Arrhythmia detection from a smartwatch, my graduation project. Galaxy Watch → Android app → FastAPI backend that reads live ECG and classifies each heartbeat into 5 classes. I compared SVM, Random Forest, XGBoost, 1D-CNN and CNN-LSTM on 109,438 MIT-BIH beats; the 1D-CNN reached 98.3% accuracy and 0.90 macro-F1.
-
MedVision · 2026
Breast ultrasound, built at a hackathon. EfficientNet-B0 classifies a scan as normal, benign or malignant; a U-Net with a ResNet-34 encoder outlines the tumour. Served as a FastAPI + React app with drag-and-drop upload and mask overlays.
-
SpinOff · 2025
Diffusion super-resolution for low-dose MRI, a computer vision course project. Reconstructs high-quality MRI from low-dose simulations on the IXI dataset, comparing bicubic, SR3 and our SR3-FiLM variant. Includes a Streamlit demo.
-
Paper Command · 2026
A game, for fun: real-time strategy drawn like doodles in the margins of a school notebook. Vanilla JavaScript with no build step, sound synthesized in the browser, and a multiplayer server.
Publications
- 2025Estimation of ERA5 tropospheric parameters using GNSS data over Tashkent
H.E. Eshkuvatov, Sh.N. Mardonov, O.V. Xudoynazarov, Z.J. Ruziev, Sh.Sh. Numonjonov, J.R. Hoshimov, F.X. Asatullayev, I.M. Egamberdiev, M.A. Musurmonov
Journal of Atmospheric and Solar-Terrestrial Physics, 106648
Smaller things
- swearguardUzbek and Russian profanity detection for chat moderation. Character n-grams and logistic regression, tuned for 0.95 precision, sub-millisecond on CPU.
- Credit defaultMerged 6 financial sources into 90K records and 85 features; LightGBM with class weighting for imbalanced default prediction.
- Emotion detectionEmotion recognition from images, video and text with OpenCV and NLP.
Experience
- 2025 – nowLead AI Engineer, SinoAI. Medical consultation chatbot on Telegram, web and mobile; medical RAG on ChromaDB and LangChain; real-time voice (VAD, STT, streaming LLM, TTS); Langfuse observability; Docker on GCP and AWS.
- nowResearch Assistant (part-time), New Uzbekistan University.
- 2025Data Science Intern, Aisha AI. Audio-text datasets for Uzbek TTS and STT.
- 2024 – 2025Project Manager and Data Analyst, CommerceHub. Ran six Uzum Market shops; grew Bikon.uz from 60M to 230M UZS a month in 3 months.
- 2024Data Science Intern, RealSoft AI. Classification and regression with scikit-learn.