Aleksandr Baranov

The task

Take a list of business names and collect directory data for them from Instagram and other sources: phone, website, name, category, city.

The problem

You have to be sure the data belongs to that exact business, or you get a scandal.

How it works

1. 1. Apify collects sources against set criteria.
2. 2. An LLM parses the bio: phone, website, name, category, city.
3. 3. Doubtful cases go out for a person to check.
4. 4. Cases where every reviewer and the LLM agree are accepted automatically.
5. 5. Cases where they disagree sharply go out for a second pass.

The review screen

1. 1. Shows a link to the source.
2. 2. Hides the other reviewers' votes and the LLM's, so it cannot prompt an answer.
3. 3. Warns where an answer disagrees with the LLM's verdict.

Cost

Spend on Apify and the LLM is capped, and a local model on Ollama does the parsing instead of a paid one.

The problem

The data for every item is out there on the internet, but looking it up one item at a time takes forever.

How it works

1. 1. The photo is recognised on the device with iOS Vision, so nothing is sent to a server to be read.
2. 2. A recognised item is looked up in the catalogue. If it is there, the data and the photo come from there; if it is not, the model or an external API is asked, and the answer is stored.

Cost

The model is only called for what the catalogue did not have. Anything found once is never asked for again.

Stack

React Native, iOS.

The task

An assistant for people who want to even out their life balance.

How it works

1. 1. The user writes a message to the assistant.
2. 2. The system reads the emotion and the intent behind it and works out the area of life: career, family, a hobby, friends.
3. 3. It looks at the current balance, what has had little time and what has had too much, and offers to schedule something.
4. 4. It looks for free slots in the calendar, allowing a sensible amount of time for that kind of activity and taking the day and the week into account.
5. 5. The user gets a proposal for a specific time, and either takes it or keeps talking.

Stack

React Native; voice or text in.

The task

Turn the free text of an enquiry (“hello, we need…”) into a CRM record with fields: who, how to reach them, what they need.

The problem

The model can miss a field or invent one, and every call to it costs money.

How it works

1. 1. An LLM reads the enquiry and pulls out the fields.
2. 2. Ordinary rule-based code checks the result and catches what the LLM dropped.
3. 3. Doubtful enquiries are held and shown to a person: which field is missing, and what record would be written.
4. 4. A checked record goes into the CRM.

Cost

A hard ceiling on model spend, and the pipeline stops when it is reached.

The task

A map of Irkutsk where you can find a place by name or category and see its hours, photos and reviews straight away.

How it works

1. 1. The open-source Leaflet engine and OpenStreetMap data, so there is no paid Google Maps key.
2. 2. A results list beside the map, with the pins clustering as you zoom out.
3. 3. Started on Firebase, rebuilt on React 19.