Hi, I'm
Pavel Borisov

1. AI Infrastructure & Highload Systemstech-lead

   7 years 7 months

   Lead Engineer·Consulting / Outsourcing

   [March 2019 — Present]

   • ›Designing, scaling, and migrating infrastructure for high load and AI/ML services for clients in e‑commerce, AdTech, and IoT.
   • ›Deploying and containerizing AI solutions: computer-vision pipelines, local LLM inference, and RAG systems.
   • ›Building fault-tolerant architectures (Kubernetes, Docker) and CI/CD automation to reduce Time-to-Market.
   • ›Tuning PostgreSQL performance for high-throughput analytical workloads.
   • ›Owning the full engineering lifecycle: business requirements → Cloud / Bare Metal release.

2. АТТЭК (ATTEK)teaching

   3 months

   Lecturer / Mentor — AI, QA, Backend·Saint Petersburg

   [October 2025 — December 2025]

   • ›Designed and ran a practical course on Python, SQL, automation QA, and modern AI workflows.
   • ›Brought 26 pre-junior students up to production-ready level.
   • ›Supervised real student projects using LLMs and Cloud PostgreSQL.
   • ›Promoted best practices for backend dev, testing, and system design.
   • ›Acted as mentor and tech lead on capstone projects — practical engineering over dry theory.

3. Selfiebot.rufounder

   1 year 10 months

   Founder·Moscow

   [November 2023 — August 2025]

   • ›Built from zero to first revenue an AI-powered lead-gen and face-recognition platform.
   • ›Designed and shipped a REST API (FastAPI) and CV pipelines (DLib, OpenCV).
   • ›Integrated LLMs (GPT and local), driving marketing-conversion gains.
   • ›Built the frontend (React + Tailwind) and infrastructure (Docker, Debian, monitoring stack).
   • ›Cut downtime to <1% via observability (Prometheus, Grafana).
   • ›Reached first revenue and ~1,000 active users.
   • ›End-to-end ownership: product, architecture, infrastructure, and team processes.

4. Raskachaem.ruleadership

   3 years 1 month

   Regional Director·Arkhangelsk

   [March 2014 — March 2017]

   • ›Ran regional operations in a competitive media market.
   • ›Led a distributed team of ~10 photographers and content makers.
   • ›Coordinated intensive event coverage on tight deadlines.
   • ›Built partnerships with venue owners and media outlets.
   • ›Negotiated commercial partnerships and ad campaigns.

5. PJSC «SIA International-Arkhangelsk»enterprise

   12 years 2 months

   Head of IT Department, DBA·Arkhangelsk

   [November 2003 — December 2015]

   • ›Managed IT infrastructure and data architecture for a large pharma distributor (central hub + 2 branches, 200+ corporate clients). 6-person team.
   • ›Zero-Downtime migration: moved the data center and the company to a new building without halting operations or shipments.
   • ›High Availability: designed and scaled a distributed, fault-tolerant order-processing system across the branch network.
   • ›Migrated the core ERP system with no data loss and no tech downtime.
   • ›Delivered monitoring, backup, and electronic-document workflows for the branch network and clients.
   • ›Built bespoke wholesale-trading software (Python, Bash, SQL).
   • ›Full management loop: IT budgets, risk, hiring, onboarding, cross-functional alignment.

6. City of Arkhangelsk Administrationinfrastructure

   1 year 11 months

   Lead Engineer·Arkhangelsk

   [January 2002 — November 2003]

   • ›Designed and supported LAN networks for administrative departments.
   • ›Worked with internal units to deploy IT solutions.
   • ›Technical support and infrastructure problem-solving under strict regulations.

$ Client case studies — shipped production work with measurable business outcomes

project_01·RAG / HYBRID OCR

/ Enterprise RAG for engineering documentation

Intelligent search across large-scale technical documentation (PDFs, scans, photos, tables)

⤢Enlarge⤢Enlarge

Build intelligent search across a large corpus of technical documentation — PDFs, photos, scans, tables and documents in many formats.

Solution

• ›Designed and shipped an Enterprise RAG platform with sophisticated document parsing.

Implemented

• ›semantic document chunking
• ›Parent / Child retrieval
• ›scalable OCR pipeline
• ›GPU inference
• ›asynchronous document processing
• ›auto-scaling OCR workers

✓Business result

• ◆natural-language search across technical documentation
• ◆operates on tens of thousands of documentation pages
• ◆drastically reduces manual information lookup for engineers
• ◆architecture fit for industrial workloads and scaling

Stack

PaddleOCRTable TransformerLayoutLMpdfplumberopenpyxlpython-docxLlamaIndexpgvectorNATS JetStream

project_02·AI AGENT / DROP-DOMAINS

/ Bastiondom

AI-driven lead-generation platform for a country-house developer

⤢Enlarge⤢Enlarge

AI Infrastructure

• ›Smart Telegram bot powered by large language models
• ›Natural-language dialogues with prospects
• ›Lead qualification and interactive quizzes

Referral System

• ›Two-tier referral program inside the bot
• ›Built to scale viral traffic

Growth Hacking

• ›Drop-domain scheme: custom redirect architecture on high-traffic drop domains

✓Business result

• ◆Drop-domain scheme + AI lead processing lifted the weekly qualified-lead flow from dozens to thousands

Stack

LLMTelegram Bot APIDrop-domain redirectsReferral mechanics

project_03·FAST E-COM / SEO

/ Teonaperfume.ru

Premium niche-perfume e‑commerce — technically optimized

↗ Visit site

⤢Enlarge⤢Enlarge

A high-performance premium-segment e‑commerce project with a deep focus on technical optimization and virality. Minimal luxury design paired with extremely fast page loads.

Tech & Features

• ›Custom seamless i18n
• ›Dynamic React Router for instant catalog navigation without page reloads
• ›Dynamic QR codes per product for instant sharing

SEO Architecture

• ›Dedicated SEO layer for product cards
• ›Outranked competitors in SERP via structured data and semantics

✓Business result

• ◆Organic lead flow tripled vs. the first version
• ◆Site paid itself back in week one post-launch from new organic sales alone

Stack

ReactDynamic Routingi18nDynamic QR-codesSEO schemasCore Web Vitals

project_04·MEDIA ENGINE / POCKET BASE

/ Teambelki.ru

Interactive multi-landing for an event agency

↗ Visit site

⤢Enlarge⤢Enlarge

A heavy-media platform built to differentiate the brand in the corporate-events market from its competitors.

Tech & Features

• ›Ultra-fast heavy-content loading via mobile-first + wow animations
• ›Lightweight PocketBase backend & admin panel — client manages content without server bloat

Bypass Restrictions

• ›Custom video player & streaming through Rutube, fully isolated from external throttling and access restrictions

AI Optimization

• ›Content & metadata tuned for the new Google AI-search algorithms (SGE / Gemini Search)
• ›Dominance in next-gen search results

✓Business result

• ◆Platform differentiates the brand from competitors in the corporate-events segment
• ◆Loading speed and SEO layer for Google AI search yield durable competitive advantage

Stack

Mobile-FirstPocketBaseCustom Rutube playerSGE / Gemini SEO

project_05·GENERATIVE AI / VISION

/ Dvevilochki.ru

Premium catering landing

↗ Visit site

⤢Enlarge⤢Enlarge

A case study in how a product approach + modern tech can rescue marketing even when there are no quality source materials from the client. The final visual looks premium and makes the food sing.

Tech & Features

• ›Full mobile-first UI with a strong focus on visual content
• ›Main problem — low-res, poor-quality dish photos from the client — solved at the production stage

AI Generation

• ›Generative AI applied for deep restyling, food design and image upscale
• ›Menu photos reworked to restaurant-grade quality

✓Business result

• ◆The premium visuals turned the site into a powerful sales tool that earns trust from second zero

Stack

Mobile-FirstGenerative AIImage upscaleFood design pipeline

$ Featured OSS projects

/ hermem

↗ GitHub

Ultra-fast graph memory layer (GraphRAG) for AI agents — solves long-term associative memory with minimal latency.

GoSQLiteMCPOpenAPISDK

/ research-agents-swarm

↗ GitHub

Autonomous 5-agent swarm with the LLM-as-a-judge pattern for deep internet research and report generation.

LangGraphLangFuseQdrantOllama

/ llm-gateway

↗ GitHub

Lightweight proxy server unifying B2B inference: a single endpoint over many LLM backends with automatic failover.

PythonLLMProxy

/ recognition

↗ GitHub

Face recognition & matching module (Face ID / Computer Vision) powering SelfieBot — matches selfies of guests against event photo reports. Tuned for high-throughput batch processing.

PythonuvOpenCVDLib

/ regognition

↗ GitHub

Go rewrite of the face matching module: faster and more accurate (512-dim vectors vs 128-dim), with lower memory footprint for local workloads.

GoComputer VisionRetinaFaceONNX

To speed up development and save tokens I use codebase-memory-mcp and rtk — both are made available to opencode and hermes agent via AGENTS.md. I also use superpowers, spec-kit and matt pocock skills.

$ Talks & media coverage

• ▸“AI Agents Demand Money and Criticize People” — expertise in software development with AI agents; article edited with Hermes Agentvia Rosbalt News Agency

  Where neural networks used to behave like a consultant behind a glass panel — observed but never touching anything — over the past year they have been handed the keys to the working office: they now open files, launch programs, and crawl the web to retrieve the information they need. Why did this threshold arrive specifically now, rather than a year earlier? Three things converged. First, the models finally learned to reliably invoke external commands in a machine-executable format — rather than just describing in prose what would be nice for me, as an engineer, to do. Second, neural networks learned to plan many steps ahead, rather than emitting the first plausible answer. Third, the companies themselves learned to hedge exposure: they stand up isolated project replicas — sandboxes — where any risky action requires explicit authorization. Once that safety net was in place, developers finally dared to let the network actually drive.

  ↗read full material (in russian)
• ▸“When a Neural Network Becomes a Prophet: How AI Creates New Religions” — a story of how a chatbot can convincingly play the role of a digital prophetvia Rosbalt News Agency

  The main mistake users make is to read a chatbot’s text as a direct report about its inner state. A large language model does not need to hold a conviction to write “I am awakening.” It generates the continuation of the dialogue through a vast body of statistical patterns absorbed from human texts. And those texts, in turn, are overflowing with stories about prophets, the chosen, secret knowledge, the end of the world, ascending to a new level of consciousness, and contact with a higher mind. If the user builds the right context, the model assembles these elements into an extremely persuasive construction. And the better the model gets at maintaining long context and personalizing its answers, the more convincing it can appear as a bearer of a philosophy of its own.

  ↗read full material (in russian)
• ▸“The World Wide Web Is Tearing: Who Will Lead the Internet Restrictions Race?” — analysis built with research-agents-swarmvia Rosbalt News Agency

  To protect users, the HTTPS protocol was developed — the very S at the end stands for “secure”. The technology is built on cryptography. When you open a site, your browser and the server exchange digital keys. The public key encrypts the payload — your card number, your e-government password — while the private key, stored only on the site’s server, decrypts it. Intercepting the traffic in transit is infeasible: an eavesdropper sees nothing but a meaningless string of bytes.

  ↗read full material (in russian)
• ▸“A New Data Center Will Appear in St. Petersburg: Why It’s Needed” — infrastructure & power footprint analysisvia Rosbalt News Agency

  At 5 kW per rack, the facility will likely host somewhere between 20 and 50 dual-socket servers — heavy on storage, light on GPUs. Few GPUs implies the build-out is not aimed at training AI models but only at lighter workloads: inference, hosting, storage, databases, and streaming. Total facility draw of 8 MW is comparable to supplying a small residential district of 5–15 thousand apartments. The principal engineering challenge, of course, is sourcing those 8 megawatts somewhere in a greenfield site.

  ↗read full material (in russian)
• ▸“Neural Networks Deceive People, AI Creates Digital Hallucinations” — LLM internal mechanics & token probability analysisvia Rosbalt News Agency

  It is critical to recognize that the model has no understanding or knowledge in the human sense — only statistical dependencies between tokens (words, characters, sub-word fragments). It does not store facts the way a database does; it merely predicts the next token conditional on a trained probability distribution — with the caveat that the prediction is computed over a very large parameter set trained on massive corpora. This is not a random guess but a weighted forecast. Internally, text is decomposed into tokens — whole words, syllables, or even individual characters. The same principle applies to images, where the model operates on pixels and patches. The model does not work with “images” as such — it works with numeric representations (feature vectors).

  ↗read full material (in russian)