Denis Kleptsov — WritingNotes on AI engineering, on-device ML and the modern web.https://kleptsov.com/How I use fine-tuning, quantization and speculative decoding to speed up a voice bothttps://kleptsov.com/blog/voice-bot-speculative-decoding/https://kleptsov.com/blog/voice-bot-speculative-decoding/Why an unquantized 12B won't run on a gaming GPU, how QLoRA and 4-bit quantization make it fit, and how a fine-tuned draft model makes it 1.7x faster, fast enough for a live voice call.Sun, 28 Jun 2026 00:00:00 GMTRunning a containerized web app on AWS Lambdahttps://kleptsov.com/blog/web-app-on-lambda/https://kleptsov.com/blog/web-app-on-lambda/How md2docx went from a VPS container to serverless on AWS for ~$0 — without changing a line of app code, plus the three sharp edges nobody warns you about.Mon, 15 Jun 2026 00:00:00 GMTGuardrails for an LLM that emails real peoplehttps://kleptsov.com/blog/guardrails-llm-email/https://kleptsov.com/blog/guardrails-llm-email/An assistant that auto-replies to email can do real damage. The boring safety layer is what makes it shippable.Fri, 12 Jun 2026 00:00:00 GMTEditing Office files without breaking themhttps://kleptsov.com/blog/ooxml-document-surgery/https://kleptsov.com/blog/ooxml-document-surgery/Word, Excel and PowerPoint are ZIP archives of XML. Here's how to change the text and keep every font, table and chart intact.Tue, 09 Jun 2026 00:00:00 GMTAn ensemble that explains itselfhttps://kleptsov.com/blog/ensemble-that-explains-itself/https://kleptsov.com/blog/ensemble-that-explains-itself/Fast local classification you can trust, plus an LLM that justifies each call with quotes — without letting it touch the verdict.Fri, 05 Jun 2026 00:00:00 GMTThe case for on-device LLMshttps://kleptsov.com/blog/case-for-on-device-llms/https://kleptsov.com/blog/case-for-on-device-llms/Cloud models are convenient until the data is confidential. Why I keep shipping apps where the model runs entirely on your machine.Tue, 02 Jun 2026 00:00:00 GMTLet the database say no: preventing double-bookinghttps://kleptsov.com/blog/database-enforced-no-double-booking/https://kleptsov.com/blog/database-enforced-no-double-booking/Two people, one slot, the same millisecond. Application-level checks lose this race — Postgres can win it.Wed, 27 May 2026 00:00:00 GMTShipping agentic RAG without the hypehttps://kleptsov.com/blog/shipping-agentic-rag/https://kleptsov.com/blog/shipping-agentic-rag/Most RAG demos fall over in production. Here is the small set of things that actually made mine reliable — and measurable.Wed, 20 May 2026 00:00:00 GMT