AI / ML engineering · Yerevan (GMT+4) · EN / RU / HY

AI systems for business, built to be checked.

Assistants that answer from your documents with cited sources, support bots, workflow automation, data pipelines and custom models. Every project starts with a written scope and a pilot on your own data, and ends with tested code, documentation and ownership on your side.

START HERE

Start here: a measured audit

AI assistant audit (RAG). A fixed-scope check of how often your document assistant is right, with numbers you can act on.

Who it is for
Teams that already have a chatbot or assistant over their documents and are unsure how often it is right.
You provide
The assistant (or access to it), a sample of the documents it answers from, and a reviewer who can judge correctness.
Format
Fixed scope, 3 to 10 working days depending on size. Three sizes:
What you get
  • An evaluation set built from your own documents and real questions; you keep it
  • Measured scores: retrieval hit rate, citation accuracy (does the cited passage support the answer), answer faithfulness, correct "I don't know" on unanswerable questions, and latency
  • A short written report with the 5 highest-impact fixes, ranked by measured effect
  • Optional: implement the fixes and re-measure
SmallAbout 25 questions.
StandardAbout 60 questions and a fixes plan.
FullAbout 120 questions, fixes implemented and re-measured.

Proof. The evaluation suite is open: doc-rag-api (see the README "Evaluation" section: 109 questions, hit@3 of 94 to 100%, citation precision and recall reported, and the weak spot stated honestly). The system it measures is live at ask-documents.

Services

Seven productized services

Each has a defined scope, a typical timeline and a stated way of checking quality. Timelines are confirmed in the written scope.

01

Document AI assistant with cited answers

Question answering over your own documents, every answer linked to its source passage.

Try it: ask documents →

Who it is for
Teams with manuals, policies, contracts, product documentation or a support knowledge base that staff or customers search by hand.
Problem
People cannot find the right paragraph in a large document set, and generic chatbots answer without evidence.
You provide
A sample of the documents (50 to 200 pages is enough for the pilot), 20 to 40 real questions people ask, a named reviewer who can judge answer correctness.
What is delivered
  • Ingestion pipeline for PDF, DOCX, HTML and plain text, with chunking and an index (keyword search, optionally embeddings)
  • FastAPI service with a documented endpoint that returns the answer plus the cited source passages
  • Optional LLM answer layer with a fallback that returns the retrieved passages only
  • Simple web or chat interface, or an API for your existing product
  • Evaluation set of real questions, with a retrieval and answer-quality report
  • Deployment notes and a short operator guide
Typical timeline
3 to 5 weeks
How quality is checked
Automated tests in CI; an evaluation set built from your questions; every answer carries citations so a reviewer can check it; refusal when no source supports an answer. Reference: doc-rag-api on GitHub (24 tests in CI).
02

Customer support and lead-qualification bot

A bot for web chat, Telegram or WhatsApp that answers common questions, qualifies enquiries and books or hands over to a person.

Who it is for
Service businesses, agencies, clinics, schools and online shops that receive repetitive enquiries in messengers.
Problem
Staff answer the same questions all day, and new enquiries wait hours for a first reply.
You provide
Your FAQ and service information, the booking or qualification rules, access to the messenger account or Business API, the target tone of voice.
What is delivered
  • Bot for one channel (web widget, Telegram or WhatsApp) with scripted flows plus LLM answers from your FAQ
  • Booking or lead capture written to Google Sheets, a CRM or e-mail
  • Human handover with the conversation history attached
  • Admin settings for opening hours, price lists and answer texts, edited without code
  • Conversation checks and a monitoring log
  • Handover document covering deployment, credentials and maintenance
Typical timeline
2 to 4 weeks
How quality is checked
Scripted test conversations covering the main flows and failure cases; answers restricted to supplied knowledge; review of real transcripts during the pilot. Reference: telegram-ai-assistant-bot (booking and FAQ, 63 tests; publication pending).
03

Workflow automation with LLMs

Repetitive office processes turned into reliable pipelines using n8n, Make or Python, with an LLM only where judgement is needed.

Who it is for
Operations, sales and finance teams that move data between e-mail, forms, spreadsheets, CRMs and messengers by hand.
Problem
Manual copy-paste work is slow and error-prone, and one-off scripts nobody maintains break silently.
You provide
A walkthrough of the current process with real examples, access to the tools involved, a person who can confirm edge cases.
What is delivered
  • Workflows built in n8n, Make or Python, depending on what you already use
  • LLM steps for classification, extraction, summarising or drafting, with structured output validation
  • Error handling, retries and alerts on failure
  • Run log so every execution can be traced
  • Exported workflow files and setup instructions
  • Short training session for the person who will operate it
Typical timeline
1 to 3 weeks
How quality is checked
Each workflow is run against a set of real example inputs before launch; LLM outputs are validated against a schema; failures raise an alert rather than passing silently. Reference: six n8n workflows with diagrams and downloadable JSON.
04

Data extraction and scraping pipelines

Scheduled collection of data from websites, documents and APIs into clean, structured tables.

Who it is for
Analysts, e-commerce and research teams that need public or licensed data in a usable format on a regular basis.
Problem
The data exists on the web or in PDFs, but not in a form that can be filtered, compared or fed into a model.
You provide
The list of sources and the fields you need, confirmation that collection from those sources is permitted for your use, destination for the data.
What is delivered
  • Collector scripts in Python with scheduling
  • Parsing and cleaning with a documented schema
  • Output to CSV, a database or Google Sheets
  • Deduplication and change tracking between runs
  • Data-quality checks and a summary report per run
  • Documentation of sources, limits and how to extend the collector
Typical timeline
1 to 3 weeks
How quality is checked
Field-level validation and row counts checked on every run; sample records compared manually with the source during the pilot; failures logged with the affected URL. Collection respects site terms and robots rules.
05

Custom machine learning models

Classification, forecasting and computer-vision models trained on your data, with measured accuracy.

Who it is for
Companies with historical data or images and a concrete prediction or detection question.
Problem
Rules and spreadsheets no longer capture the pattern, and there is no one in-house to train, evaluate and deploy a model.
You provide
The dataset or access to it, a definition of what a correct prediction is, labels or a way to obtain them, a domain contact for questions.
What is delivered
  • Data audit and baseline model, so any improvement is measurable
  • Trained model with a reproducible training script
  • Evaluation on a held-out set with metrics suited to the problem and an error analysis
  • Inference API or batch script
  • Model card describing data, limits and known failure cases
  • Retraining instructions
Typical timeline
3 to 6 weeks
How quality is checked
Held-out evaluation and comparison against a simple baseline; fixed random seeds and versioned code; honest reporting of limits. Reference: stereo-distance-calculator (6% median distance error on 27 measured points) and wardrobe-intelligence-lab (computer vision with a local LLM).
06

AI feature for your product

An LLM or ML capability added to an existing application through a clean API, with tests and evaluation.

Who it is for
Product teams and founders who want a specific AI feature (search, summarisation, extraction, recommendations, assistants) inside an app they already run.
Problem
Prototypes work in a notebook but are not tested, measured or safe to ship inside a production application.
You provide
Repository or API access, a description of the feature and its success criteria, a technical contact on your side.
What is delivered
  • Feature implemented in your stack or as a separate FastAPI service
  • Prompt and model configuration kept in code, with provider fallbacks
  • Automated tests and an offline evaluation set for the feature
  • Cost and latency measurement
  • API documentation and integration examples
  • Code review notes for your engineers
Typical timeline
2 to 5 weeks
How quality is checked
Unit and integration tests in CI, typed code and linting, regression evaluation whenever a prompt or model changes. Reference: silent-stakeholder (250+ tests, mypy and ruff in CI) and resume-constructor (React, TypeScript, Supabase, Playwright and accessibility checks).
07

White-label engineering for agencies

AI and automation delivery behind your brand, for agencies that sell the project but need an engineer to build it.

Who it is for
Web, marketing and software agencies whose clients ask for chatbots, automations or AI features the agency cannot yet build.
Problem
Declining the work loses the client; hiring a full-time specialist is not justified yet.
You provide
The client brief, one point of contact at the agency, access to the client environments through your accounts.
What is delivered
  • Scoped work package built to your client brief
  • Delivery under your name, with no mention of my involvement unless you want it
  • Repositories transferred to your organisation
  • Client-ready documentation written for your branding
  • Support during the client acceptance phase
Typical timeline
Set per package
How quality is checked
Same standard as direct projects: written scope, tests, documented handover. NDA signed before the brief is shared.

Engagement models

Choose how to start

Fixed-scope project

One written scope, one deliverable list, agreed timeline. Suited to a well-defined need.

Pilot, then build

A short pilot on your own data proves the approach before the full build is commissioned. Suited to uncertain scope.

Monthly support retainer

Reserved hours for maintenance, monitoring, improvements and questions after handover.

White-label or subcontract

Work delivered behind an agency or a team, under their brand and process.

Process

From first call to handover

01

Discovery call

30 to 45 minutes to understand the problem, the data and the people involved. No obligation.

02

Written scope

A short document: goal, deliverables, what is out of scope, timeline, what you provide, how success is measured.

03

Pilot on your data

A first version running on a sample of your real material, so decisions are based on evidence.

04

Build

Iterations with visible progress, tests and evaluation throughout.

05

Handover and support

Code, documentation and a walkthrough; a support period for questions and fixes.

Working terms

What you can rely on

NDA on request

Signed before you share confidential material.

You own the result

Code and deliverables transfer to you on completion of the agreed work.

Documentation

Every project ends with setup, operation and maintenance notes.

Written scope

Work starts from an agreed document, and changes are agreed in writing.

Working hours

Based in Yerevan (GMT+4). Languages: English, Russian, Armenian. Daily overlap with European and Moscow working hours.

Data handling

Your data is used only for your project. Access uses accounts and keys you control; credentials are never stored in code.

Public work

Code you can inspect

Public repositories on GitHub, with their tests and CI. A Telegram booking and FAQ assistant (63 tests) and six n8n automation workflows are published with diagrams and downloadable JSON.

  • doc-rag-api
    FastAPI document Q&A with cited sources; BM25 plus optional embeddings or LLM; 24 tests in CI
  • silent-stakeholder
    Evidence-grounded product intelligence from app reviews and roadmap; 250+ tests; mypy and ruff in CI
  • wardrobe-intelligence-lab
    Computer vision with a local LLM stylist; live Streamlit application
  • resume-constructor
    AI resume builder in React, TypeScript and Supabase; Playwright and WCAG checks
  • think-weather-cli
    Command-line tool with 340 offline tests and CI on three operating systems
  • stereo-distance-calculator
    Stereo-vision distance estimation; 6% median error on 27 measured points
  • batchcor-rna-embeds
    scGPT embeddings, batch correction and survival prediction on RNA data

Final-year Informatics and Applied Mathematics, Yerevan State University. ML Engineer internship at Mathrix AI (spring 2026). Advanced Machine Learning program, Armenian Code Academy.

Contact

Start with a 30-minute call

Describe the problem and the data you have. You receive a written scope; NDA on request.