neudahin.com

AI agents · harnesses · system design

AI agents and AI engineering for business

I design and build AI agents for business processes — in the channels where the work already happens: web, messengers, voice, and integrations with your CRM, ERP and telephony. I am responsible for their reliability: tooling and permissions, control over actions, quality evaluation, and regression when a model changes. The agent prepares an action; a human approves it. I work remotely, under contract.

Discuss a task Message on Telegram

Александр Неудахин

15+

years in software development, self-employed since 2011

3

my own AI platforms — each registered as a computer program, 2026

6

model categories in my own BiFL platform — from chat and embeddings to video generation

PhD

in engineering — neural-network knowledge representation models

Expertise

Six areas where I own the task end to end — from an agent built around your process to the system that runs it in production.

  1. AI agents for business processes

    The agent parses the intent and prepares an action; a human approves it, and mutating operations are protected by a preview. It works where the process already lives — web, messengers, voice — and plugs into your CRM, ERP and telephony.

  2. Harnesses and agent reliability

    The engineering around the model: tools and permissions, context assembly, scenario sets for quality evaluation, regression when a model changes, cost per request. External agents connect to the system over MCP.

  3. Automating routine operations with LLMs

    Handling inbound requests, documents, and data moved between systems. I draw the automation boundary — what goes to the machine and what stays with people — and take it to a measurable effect.

  4. System design and architecture audit for large web systems

    Load, queues, data, fault tolerance. An audit before growth — to see what will hit the ceiling first — or design from scratch: from the data schema to operations.

  5. RAG and search across company knowledge

    Text extraction from PDF, DOCX, XLS and CSV, four chunking strategies, vector search in Qdrant, local FastEmbed embeddings with no external APIs.

  6. Production engineering

    Async pipelines on FastAPI and asyncio, queues and webhooks, billing with subscriptions and pay-as-you-go, Docker and CI/CD — everything that turns a demo into a service.

Ways to work together

Four ways to work together — from a full project to support for a system already in production. Each is described here, so the format can be picked without a round of emails.

  • End-to-end project

    The whole task: analysis, a prototype, the system in production, launch. The result is a working service.

  • Architecture and audit

    A review of a system or a design: load, queues, data, fault tolerance. The result is a diagram, the risks and a plan of changes.

  • Long-term work

    A long contract with your team — as a sole trader, remotely. The technical side of the project is on me.

  • Ongoing support

    The system after launch: agent quality, model updates, evaluation scenarios, incidents.

Work starts with a review of the task: the outcome is a format, a plan and an estimate.

Foundation

Research underpins the engineering work — it is not the main subject of this site: the short version is here, the full lists are in the archives.

  • PhD in engineering (2009) — neural-network knowledge representation models.
  • 14 research publications, 2008–2026: articles, conference papers and a dissertation. Publications
  • 4 computer-program registration certificates, three of them sole-authored, 2026. Certificates
  • 3 grant projects with the Foundation for Assistance to Innovation, one of them an application under review. Grants

Tech stack

The technologies behind the case studies — from neural models to infrastructure.

Message me

Pick a subject and tell me what you need. You will get a to-the-point reply — no twenty-question brief.

Or message me on Telegram for a faster reply.

The subject hints at what to include: a budget and timeline for a project, a schema or system description for an audit, the expected workload for long-term work.

For an audit, attach a schema or a system description — we will start from facts.

Optional. Most often relevant for an end-to-end project and long-term work.

Optional. Most often relevant for an end-to-end project and long-term work.

Optional. One file up to 10 MB — PDF, document, spreadsheet, archive or image.

Consent Policy

Need a CMS website or an ERP integration? That is the work of my studio makeWWW (in Russian).