01
Applied AI
LLM systems inside a workflow.
Retrieval, agents and multimodal prototypes, built around a decision someone makes every day. Where an action has consequences, the model proposes and a person confirms.
Read the cases →Profile / 01AI systems · research · notes
AI / ML Engineer
Fedor
Pakhurov
I build AI systems around real workflows.
Most of my work starts with a process, not a model: something is slow, manual or hard to scale. I use LLMs, retrieval and agents where they actually improve it, and take the result from a prototype to something measurable that runs in production.
02 / Focus
Three sides of the same work.
The model is usually only one part of it.
02
Systems
Serving, latency and everything around the model.
Local inference, quantization, integrations and telemetry. A prototype is not finished until it is fast enough, observable and allowed to touch the data it needs.
Engineering projects →03
Research
Not only whether it works, but why.
Evaluating diffusion-model errors and testing wiring rules from C. elegans connectomes.
Research →03 / In practice
Three examples, with the numbers.
From production work. Employer, system and client names are left out.
Retrieval
In the main retrieval workflow, getting a usable answer took about 4 minutes per request. After we rebuilt retrieval around the decision people actually had to make and tuned inference for latency, it took about 5 seconds.
Read the case →Local inference
A Text-to-SQL assistant had to move inside the company perimeter onto one A100, with a target of under 10 seconds per question. On transformers it needed 2 minutes; on vLLM with FP8 quantization, 3 seconds.
Read the case →Agents
Supply, transit and export agents under one coordinator, with a person confirming every critical operation. The platform handles 5,000+ operational actions a month, with an estimated annual effect above ₽48M.
Read the case →
04 / Systems
AI transformation is a chain, not a feature.
Foundation
Capture and structure decisions.
- Paper
- Digitalization
- Decision history
- Business rules
Delivery
Train, operate and support models.
- Dataset
- ML
- MLOps
- Infrastructure
Agency
Connect models to action.
- LLM
- Agents
05 / Knowledge
Notes built to be studied, not scrolled.
Course notes for Generative Models and NLP, taught at the HSE Faculty of Computer Science. Prerequisites, equations, permanent URLs and a clear progression.