svezhest

Jan 2024 - Sep 2025. Core engineer on a market-leading AI coding assistant (60k monthly active users). - Prototyped a project-wide RAG retrieval layer (client-side IDF-scoring over project files, open tabs, and the LSP graph) that lifted next-line prediction accuracy from 70% to 85%; A/B testing surfaced a latency cost on low-end machines, so we held it back. - Designed the inference token-aware, shared-state load balancer, cutting end-to-end inline-suggestion p95 latency from 1-2s to under 500ms. - Owned every new model integration end-to-end (quantization, inference engine, context size, tool-calling parsers) and the regexp-based response-guidelines filter. - Built the backend for the assistant's agent mode (/v1/compactions, /v1/responses) and an automated MLflow pipeline (RepoBench, M2RC-Eval, SAFIM) to benchmark inference frameworks (vLLM, SGLang) across models.

Jul 2023 - Oct 2023. Worked with the team responsible for algorithmic bidding strategies driving a major share of ad revenue; shipped period-based budgeting analysis that eliminated rounding discrepancies in spend allocation.

Feb 2023 - Jun 2023. Solo built an image-composer MVP using Stable Diffusion XL with IP-Adapter.

Nov 2023 - Jan 2024. Solo implemented Yandex Speech API and a local Whisper Large-v3 fallback ASR engine over WebSocket, enabling voice control of a Small-Angle X-ray Scattering (SAXS) instrument.

Sep 2022 - Jan 2023. Built a Telegram "Random Coffee" bot with a smart user-matching queue and meet-up suggestions, working in a team of 3.