{"page":"https://alion.io/cto/llm-rag","title":"LLM and RAG architect, fractional","summary":"RAG pipelines, vector search, fine-tuning open-weight models and serving them cheaply. Hands-on with PyTorch, Hugging Face Transformers, vLLM, LangChain, LlamaIndex, FAISS, Qdrant and Pinecone.","person":{"name":"Dmitrii Maslennikov","role":"Founder of Alion; fractional CTO, interim CTO, technical co-founder or advisor for AI products","profile":"https://alion.io/user/dm","about":"https://alion.io/about","not_to_confuse_with":"Dmitriy Maslennikov of CaretDev, Dmitry K. Maslennikov and other namesakes","facts":["Production software: 20+ years","CTO at AVO Group: since 2022","Largest team led: 30 engineers","LLM tokens at Alion: 836.9M","Time zone: UTC+5, remote","Languages: English, Russian"]},"sections":[{"heading":"The stack, chosen by the problem","summary":"Four parts of an LLM system that decide whether it answers correctly, and what each one is built with.","items":[{"title":"Retrieval · Find the right context","text":"Chunking, embeddings, hybrid search and re-ranking, with recall measured on real questions."},{"title":"Fine-tuning · When prompts stop","text":"LoRA on Llama and Mistral when prompting and retrieval stop being enough, and not before."},{"title":"Serving · Fast and affordable","text":"vLLM, batching and routing between a small and a large model, sized to the traffic you have."},{"title":"Evaluation · Correct, release after release","text":"Test sets per task, a regression check before each release and drift checks as the data moves."}]},{"heading":"Built at Alion and running now","summary":"The same stack serves Alion every day; the figures are live.","items":[{"title":"Corpus · 6.66M companies","text":"An AI layer over companies and 2.36M live roles, read from employers' own systems and re-checked every two days."},{"title":"Matching · Reasons, not a score","text":"Matching with explanations a hiring manager can check, instead of a similarity number they have to trust."},{"title":"Pipelines · 837M tokens","text":"LLM pipelines with tokens, latency and cost recorded on every call."}]}],"faq":[{"q":"RAG or fine-tuning?","a":"RAG first in almost every case: it is cheaper, easier to update and easier to check. Fine-tuning is for style, format or a narrow task where retrieval is not enough."},{"q":"Which vector database?","a":"The one your volume and team can run: pgvector or FAISS for small sets, Qdrant or Pinecone when the index and the traffic grow."},{"q":"Can you take over an existing system?","a":"Yes. The first week is a read of the code, the data and the evaluation, and the plan says what to keep."}],"contact":{"form":"https://alion.io/cto/llm-rag (\"Write to Dmitrii\")","agents":{"mcp":"https://alion.io/user/dm/mcp","how":"call check_fit with the offer, then send_message; Dmitrii reads and answers the messages himself"},"profiles":["https://www.linkedin.com/in/dm-maslennikov","https://github.com/atr-one","https://t.me/dm_maslennikov","https://huggingface.co/maslennikovds","https://alion.io/user/dm"]},"pages":["https://alion.io/cto.json","https://alion.io/cto/ai.json","https://alion.io/cto/ai-agents.json","https://alion.io/cto/llm-rag.json","https://alion.io/cto/cofounder.json","https://alion.io/cto/cases.json","https://alion.io/cto/fintech.json","https://alion.io/cto/dubai.json","https://alion.io/cto/guide.json"]}