Available now · Yerevan (GMT+4)

Gevorg Hovhannisyan

AI / ML Engineer · Yerevan

I build AI assistants that answer from your own documents with sources, automations that remove repetitive work, and ML models trained on your data.

A working version firstEN / RU / HYAI video: HY · RU
Stereo-vision distance measurement: two cameras triangulate a target; predicted versus tape-measured distance for 27 real points, median error 6%
Two webcams, no neural network: 6% median error on 27 real points.Project ↗

Numbers that are true

  • 6%median error, stereo-vision distance27 tape-measured points
  • 58tests run in CI on doc-rag-apipytest, GitHub Actions
  • 340offline tests in Think WeatherCI on 3 operating systems
  • 250+tests in Silent Stakeholdermypy and ruff in CI
  • 26cited passages, all checked against sourcesSilent Stakeholder test set
  • 3languages: Armenian, Russian, Englishassistants and bots in all three

What I can build for you

Six things I build for businesses and teams.

A working version on your own data first, then you decide.

01

Support assistant over your docs

Answers from your own documents with citations, and hands over to a human when it isn't sure.

Outcome: customers get an answer with its source; unclear questions go to a person, not a guess.

02

Lead qualification and booking

n8n or Make workflows that capture a lead, ask the right questions, and book the call.

Outcome: every lead gets the right questions and a booked slot, with no manual chasing.

03

Telegram and WhatsApp bots

Bots for orders, bookings, FAQs and alerts, in English, Russian or Armenian.

Outcome: orders, bookings and FAQs handled in the app customers already use, in their language.

04

Scrapers and dashboards

Scheduled collection from public sources into Sheets or a dashboard, with retries and alerts.

Outcome: fresh data lands on schedule; a failure sends an alert instead of going silent.

05

ML models on your data

Churn, risk and forecasting models, evaluated honestly on data the model has not seen.

Outcome: a prediction you can judge, scored on unseen data, with its limits written down.

Selected work

Real projects, what they solve, how they are built.

Labels show what you can actually open: Live means a running service, Code means a public repository.

Stereo distance measurement from two webcams: predicted versus tape-measured distance for 27 real points
CodeResearch

Stereo-vision distance from first principles

Problem
Measure distance with two ordinary webcams, no neural network.
Built
Every step by hand: lens model per camera, matching the same point in both images, triangulation. 6% median error on 27 tape-measured distances (0.4 to 4.7 m), each predicted with that point left out of calibration.

Python · NumPy · ZNCC matching

doc-rag-api answering a question with cited passages
Open source

doc-rag-api

Problem
Answers from documents are useless if you cannot check them.
Built
Document Q&A: each answer comes with the exact passages it was built from. If the documents don't cover a question, it says so. Runs offline on keyword search, with optional embeddings and a local LLM via Ollama.

Python · FastAPI · RAG · Ollama

Open source

telegram-ai-assistant-bot

Problem
Small businesses lose customers when messages wait overnight and booking means a phone call.
Built
A Telegram assistant that answers from the business's own FAQ in Armenian, Russian and English, books visits with no double-booking, sends 24 h and 2 h reminders with a confirm button, lets customers cancel or reschedule, and gives the owner a daily digest. 110 automated tests.

Python · aiogram 3 · SQLite · BM25 · Docker

Open source

n8n-ai-automation-templates

Problem
Teams want AI automations but start every workflow from a blank canvas.
Built
Ten import-ready n8n workflows: lead qualification, Gmail triage, Telegram FAQ, bookings, news monitor, review analysis, real-estate lead matcher, WhatsApp AI receptionist, reels content factory and review replies. Each one is checked by a validator in CI.

n8n · OpenAI · Google Sheets · Telegram · WhatsApp

AIrealty: real-estate requests matched with owner listings
Live · client productionPrivate codebase

AIrealty: new clients every day for real-estate agencies

It never sleeps, so no buyer waits for an agent to catch up.

Problem
People in Yerevan look for flats in three languages across many places, and agents answer too late.
Built
A system that brings every request into one place around the clock, understands it in Armenian, Russian or English, and sends the agent a ready card with budget, district and best matching homes. Found 1,103 people looking for housing in one week; search across 46,000 homes answers in under a second.

Python 3.12 · FastAPI + HTMX · TypeSafe Jev AI · Telegram

Silent Stakeholder review analysis screen
Live buildCode

Silent Stakeholder

Problem
Product teams drown in app-store reviews.
Built
A pipeline that reads thousands of reviews next to the public roadmap and shows which real user needs the roadmap misses. All 26 cited passages in its test set were verified against the source reviews.

Python · LLM pipeline · Streamlit

UMAP of single-cell RNA embeddings after batch correction
CodeBostonGene program

Batch correction for single-cell RNA

Problem
Cancer RNA data from different hospitals does not line up; each lab measures a little differently.
Built
A neural autoencoder that removes the lab effect while keeping the biology. 2,218 kidney-cancer samples from 3 cohorts in one shared space.

Python · scGPT embeddings · autoencoder

Wardrobe Intelligence Lab outfit recommendation screen
Live buildCode

Wardrobe Intelligence Lab

Problem
Deciding what to wear from clothes you already own.
Built
A recommender that picks today's outfit from weather, occasion and style, and learns from likes and dislikes. The stylist chat runs on a small open model on an ordinary CPU, no paid API, in EN/HY/RU.

Python · local LLM · recommender

Sentinel crypto research dashboard
Running 24/7Paper trading only

Sentinel, a crypto research engine

Problem
New tokens hide the usual traps: frozen liquidity, a dev who can mint more, tokens you can buy but not sell.
Built
An engine that watches new tokens on Solana, Base and BSC and checks each one before anyone looks. It follows every token for 30 days and reports what would have happened after fees, so the numbers are measured, not promised.

Python · TimescaleDB · risk engine · alerting

EdgeFinder FX signal bot
Live Telegram bot

EdgeFinder, an FX signal bot

Problem
Forex traders act on noise and hope.
Built
A Telegram bot that weighs trend, session, news calendar, volatility and spread across 11 pairs, then names one setup or says stay out. Two years of data, 18 months of walk-forward tests and a sealed 3-month test found no proven edge, so it says NO TRADE instead of selling hope.

Python · LightGBM · Telegram · kill switches

Cadence automation control room
Live build

Cadence, an automation control room

Problem
Data jobs must run unattended and tell you when they fail.
Built
Scheduled collection from public sources, automatic retries, a log of every run and alerts. 1,247 rows delivered to Google Sheets on schedule.

Scraping · cron scheduling · alerts

Aurel online storefront
CodeLive build

Aurel storefront

Problem
Small shops need a fast store that feels native on a phone.
Built
Product grid, product pages and cart in plain HTML and CSS, no page builder. The base I use for shop and landing-page projects.

HTML · CSS · responsive

Think Weather terminal dashboard
Open source

Think Weather

Problem
Quick weather without leaving the terminal.
Built
Current conditions, an hourly chart and a 5-day forecast for any city or postal code, or plain JSON for scripts. No API key needed: free Open-Meteo data with caching.

Python · Rich CLI

Resume Constructor home screen
Live buildCode

Resume Constructor

Problem
Hard to know how well a CV fits a job.
Built
Upload a PDF or Word CV; it scores it against 36 roles or a pasted job ad, suggests stronger wording and exports an ATS-readable file. It never invents experience: suggestions are checked against the original resume.

React + TypeScript · Supabase · Gemini

Example sites for local businesses

BrightSmile Dental example website
Example site

BrightSmile Dental

Built
Booking-first site for a dental clinic: online appointment calendar and a 24/7 bilingual assistant. Mobile-first.
Lumière Beauty Studio example website
Example site

Lumière Beauty Studio

Built
Booking-first site for a beauty salon: service menu, time-slot selection, gallery and direct WhatsApp booking.

How a project runs with me

Four steps, and you see it working before you pay.

  1. A call

    5 minutes in chat: what you want automated, collected or answered.

  2. A working version

    A working version within 24 hours, on your own real data, before you pay anything.

  3. Build with tests

    You agree, I build it with tests, and it goes live in 1–3 days.

  4. Handoff and support

    Code, setup guide and handoff document, plus a full month of free changes after launch.

How I work

Show it working, then agree on it.

A working version first

You see the thing running on your own material before you commit to anything.

Cites sources, admits gaps

Assistants point to the passage behind every answer and say "I don't know" when the documents don't cover it.

Tests and handoff docs

You get tests, setup notes and a handoff document, so the system keeps working without me.

English C1Russian C1Armenian native

FAQ

Short answers.

Do you work remotely?

Yes. I work from Yerevan (GMT+4) and everything runs over chat and calls, so location is not a limit.

Which languages can we work in?

English, Russian or Armenian. Assistants and bots can answer your customers in all three.

How do you handle our data and privacy?

Assistants answer only from the content you approve. Nothing is invented: if your documents don't cover a question, the assistant says so and hands the conversation to a person.

What happens after launch?

You get the code, tests, a setup guide and a handoff document, plus a month of free changes after go-live.