Denis Kleptsov
On-device LLM apps, production RAG & speech-analytics platforms, and the full-stack products that wrap them.
I'm an AI/ML engineer who turns research-grade models into products people use every day — offline LLM apps for macOS, production RAG and speech-analytics systems, and the full-stack web that ships around them. I care about latency, privacy, and shipping things that work.
Shipped and running - try them.
Translate Word, Excel & PowerPoint fully offline — on your Mac.
A macOS app that translates Office documents with a 1.8B-parameter LLM running locally on Apple Silicon. Every font, table and chart is preserved, and nothing ever leaves your machine.
A self-hosted Calendly alternative that never double-books.
A scheduling app with timezone-perfect availability and database-level double-booking prevention. Yes — the same one this site uses for the "Book a meeting" button.
Turn Markdown into a beautifully formatted Word document in one paste.
A live web tool that converts Markdown into clean, styled .docx files — headings, tables, code blocks and themes included — with no sign-up.
Agents, RAG and NLP experiments on GitHub.
A production-style RAG agent, traced and evaluated end-to-end.
A retrieval-augmented assistant built on LangGraph, with full tracing and evaluation through LangSmith and a FastAPI + React surface.
An agent that writes Python, runs it, and fixes its own mistakes.
An autonomous coding agent that generates Python, executes it, catches its own runtime errors, and self-corrects in a loop until the task passes.
A clean, hackable Agentic RAG question-answering system.
A minimal LangChain example of agentic retrieval-augmented question answering — a good starting point to learn from or fork.
Production AI and ML systems shipped for enterprise clients.
Designed and shipped retrieval platforms — combining knowledge graphs with embedding search — that power internal AI assistants across an enterprise, with evaluation pipelines that keep answers honest.
Built a neural-network churn-prediction system for a mobile carrier with 50M+ subscribers and led the team through to production. Risk scoring projected to cut churn by 5%.
Stood up a secure, self-hosted code-execution sandbox so AI assistants can run generated code safely — an in-house alternative to hosted sandboxes, with isolation and resource limits.
Trained and deployed an ensemble classifier (TF-IDF + embeddings + logistic regression) over 32k+ labeled examples, with an LLM generating cited justifications for every prediction — experiments documented from data size to final model choice.
Built a production voice agent for a law firm handling inbound and outbound calls: SIP/WebRTC intake dialogues that extract structured client data, plus an outbound campaign engine with compliance windows and retry logic. About 65% of calls run end to end without a human operator.
Built a full-stack app that turns PDF and PPTX decks into interactive AI lectures: GPT narration, local TTS voice-over, real-time voice Q&A with Whisper, and auto-generated quizzes. Lecture prep dropped from about two days to under ten minutes.
Built a pipeline that reads incoming RFQ emails with PDF and Excel attachments through OCR, checks inventory, sources supplier offers over API, and drafts the commercial proposal. Up to 5x faster than the manual workflow.
Built the NLP module for a national railway's innovation-management platform: categorization, fuzzy search, and analytics across 40,000+ employee innovation proposals.
Built an internal platform that turns training documents into multiple-choice assessments with LLM-generated questions and explanations, auto-assigns them across offices and roles, and reports pass-rate analytics in real time.
Impact and stack only — client details under NDA.
What I reach for, grouped by layer.
RAG / GraphRAG · LangGraph agents · LangSmith eval · OpenAI · Embeddings · llama.cpp · Apple MLX · Guardrails
scikit-learn · Ensembles · Fine-tuning · Qdrant · TF-IDF · Classification
Python · FastAPI · SQLAlchemy · Node.js · PostgreSQL · Redis · arq
TypeScript · React · SvelteKit · Astro · Tailwind v4
Tauri · Rust · Electron · macOS / App Store
Docker · Traefik · GitLab CI/CD · Linux / VPS · Let's Encrypt
AWS Lambda · Lambda Web Adapter · Function URLs · CloudFront · ECR · ACM · IAM · CloudWatch · Graviton / arm64
Notes on building with LLMs, on-device ML and the web.
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.
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.
An assistant that auto-replies to email can do real damage. The boring safety layer is what makes it shippable.
Have an AI product to build, or a hard ML problem? Send a note or book a slot.