Alion · Fractional CTO · LLM and RAG architecture

Retrieval and models that answer correctly at scale.

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.

Dmitrii Maslennikov
Dmitrii MaslennikovFounder of Alion · fractional CTO for AI products
At a glanceUpdated 22 Nov 2026
Production software20+ years
CTO at AVO Groupsince 2022
Largest team led30 engineers
LLM tokens at Alion836.7M
Time zoneUTC+5, remote
LanguagesEnglish, Russian
Agents: alion.io/user/dm/mcp
The stack

The stack,chosen by the problem

Four parts of an LLM system that decide whether it answers correctly, and what each one is built with.

Retrieval

Find the right context

Chunking, embeddings, hybrid search and re-ranking, with recall measured on real questions.

Fine-tuning

When prompts stop

LoRA on Llama and Mistral when prompting and retrieval stop being enough, and not before.

Serving

Fast and affordable

vLLM, batching and routing between a small and a large model, sized to the traffic you have.

Evaluation

Correct, release after release

Test sets per task, a regression check before each release and drift checks as the data moves.

At Alion

Built at Alionand running now

The same stack serves Alion every day; the figures are live.

Corpus

6.66M companies

An AI layer over companies and 2.36M live roles, read from employers' own systems and re-checked every two days.

Matching

Reasons, not a score

Matching with explanations a hiring manager can check, instead of a similarity number they have to trust.

Pipelines

837M tokens

LLM pipelines with tokens, latency and cost recorded on every call.

Questions

Asked beforethe first call

RAG or fine-tuning?

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.

Which vector database?

The one your volume and team can run: pgvector or FAISS for small sets, Qdrant or Pinecone when the index and the traffic grow.

Can you take over an existing system?

Yes. The first week is a read of the code, the data and the evaluation, and the plan says what to keep.

Tell Dmitrii what you are building.

One message is enough: the product, the stage and what is stuck. He answers himself, usually within a day, and says plainly whether he is the right person for it.

An AI agent can do the same: check_fit with the offer, then send_message at alion.io/user/dm/mcp.