yomorigi

Own a multi-tenant model platform of 200+ endpoints that other teams build features on, covering model selection, embedding storage and retrieval, and serving configuration. Inference cost and latency: Profiled the whole serving path before changing anything. The cost was spread across preprocessing and postprocessing, batch scheduling, and the forward pass itself. Moved postprocessing onto the GPU in custom CUDA kernels, wrote a TensorRT plugin for an operation the framework did not support, and removed host-to-device copies by reusing buffers across requests. Reworked batching and GPU utilisation, which turned out to be the largest single win, ahead of the low-level kernel work. Benchmarked FP16 and INT8 quantization and knowledge distillation against measured latency and throughput budgets, keeping the configurations that held accuracy. Production inference cost came down roughly 100x. Models and evaluation: Built the team's evaluation stack: eval datasets and quality metrics, LLM-as-judge scoring, A/B tests on live traffic, champion-challenger against production baselines, and regression checks that gate releases in CI. Fine-tuned OCR and BERT models. Trained a CycleGAN for synthetic domain adaptation that raised detector mAP about 10% with no new annotation. Platform: Built the data layer: high-volume event streams into ClickHouse, multi-tenant dbt models, Spark and Airflow pipelines, FastAPI services, and CI/CD on Kubernetes with monitoring, alerting and automated rollback. Built an internal annotation tool several times faster than CVAT and Label Studio, and led the annotation team.

Owned marketplace search ranking and relevance, and built recommendations targeted to the audience of each product card. Found that high-selling products were missing from recommendations because their descriptions did not match what the ranking algorithm keyed on. Shipped the fix as category-level analytics feeding a RAG pipeline over pgvector, generating product-card content for 5,000+ business clients. Click-through went up 20% on live traffic. Ran comparative fine-tuning on open-source LLMs across hyperparameter configurations and benchmarked them against commercial APIs. Designed per-category validation of generated text on length, attribute ordering and keyword coverage, so nothing broken reached a live listing. Architected a microservice ERP on FastAPI, Kafka and WebSockets. Real-time updates for 1,500+ users, latency down 50%. Mentored 3 developers through code review.

Full-time to Feb 2024, part-time after. Fine-tuned OCR models for bank-receipt processing: entity identification, extraction and attribute verification on noisy, low-quality documents. Fine-tuned BERT for analytics-driven structured prediction, from data preparation through evaluation. Built fraud, risk and anomaly-detection models on transactional data. Deployed and served open-source and proprietary LLMs and image-generation models via API, including inference optimization and API-level scaling. Took a crypto exchange from scratch to production in 4 weeks with Binance API and payment-gateway integrations handling 1,000+ daily transactions. Directed a team of 3. Trained an early YOLO model for a workplace safety-compliance system, trained on AWS and served on-premise.

Built a FastAPI authorization service with JWT and CRM synchronization, managing 2,000+ user profiles at zero downtime. Developed a Telegram task-management bot with Redis caching, cutting processing time 25%. Integrated Stripe and Kassa for subscription automation, lifting revenue 15%.