Here are my production ML projects: from generative AI and recommendations to MLOps infrastructures and observability. In each case - problem, architecture, and real result.
On-prem voice AI operator handles 72% of calls without human in 0.96s with 58% cost reduction.
Problem: 600 seats in contact center, 9 min wait, SLA penalties and new AI Act requirements, regulations outdated faster than operators can learn.
Solution: On-prem stack with streaming, model cascade, orchestration, and knowledge base. Safety rules and manual escalation.
Internal tool for profiling latency, throughput, and $/req of models in production
Problem: No unified monitoring standard: teams deployed models randomly, GPUs idle, latency fluctuated, costs not tracked.
Solution: Built platform with Prometheus, Kubecost, and Torch/ONNX profiling - now visible latency, throughput, load, and $/req at model level.
MVP chat search with deployment automation, experiments, and quality monitoring
Problem: 30% of queries return no results, p95 latency 1.5s, growing cloud LLM costs
Solution: Hybrid search (vector + BM25) with fine-tuned Mistral-7B, autoscaling in K8s, and cheap inference on 1 GPU
Fraud detection system reduces inefficient spending by 24% and automates verification of 100 channels in 12 minutes
Problem: 30% of ad budget lost on channels with fraud, manual verification of 100 channels takes 25 hours
Solution: Hybrid detection system (rule-based + anomaly) with batch processing and adaptive thresholds by topics
Multimodal search and recommendation platform with full CI/CD pipeline, monitoring, and A/B experiments
Problem: High cost and p95 > 500 ms when searching 10M+ products in e-commerce
Solution: End-to-end architecture with CLIP encoder, HNSW (FAISS), TensorRT optimization, and canary A/B deployment