c. 40,000 BCCave walls

Ideas in production.

Every era moved with its builders. I’m Vitalii Kindrakevych, a senior full-stack engineer: I design and build platforms, online stores and AI systems that people use every day.

Symfony
Python
Next.js
React
Elasticsearch
Drupal
Shopify
MapLibre
Claude

I turn millions of raw records into products people use every day: depth soundings, product catalogs and call recordings become maps, online stores, search and AI tools.

Karpaty.Rocks

A guide to the Carpathians

Routes, peaks, springs and stays of the Carpathians on one map — with search that understands plain language. Below is one real route and what each technology does under the hood.

  • Symfony
  • PostgreSQL + PostGIS
  • Next.js
  • Elasticsearch
  • GraphHopper
  • Embeddings
Open karpaty.rocks

Route · elevation profile

Zaroslyak → Hoverla

01/ 06

PostgreSQL + PostGIS

The route is stored as a real geometry — 215 points with coordinates. The database measures distances and finds what is nearby.

215 points · 6,011 m · start: Zaroslyak, 1,352 m
02/ 06

GraphHopper

Builds the walking route along trails from Zaroslyak to the summit and estimates the time on foot.

6.0 km · ≈ 1 h 12 min on foot
03/ 06

Elasticsearch + Embeddings

“What is nearby” searches by meaning, not by words. The Ukrainian query «високогірне озеро біля Говерли» (high-mountain lake near Hoverla) → Lake Nesamovyte, 1,750 m.

1,024-dimension vector · 106 ms + 5 ms · nearby: Pozhyzhevska meadow, 1,500 m
04/ 06

Symfony

The API gives the app everything for this page: the geometry, the elevation profile and the places along the trail.

nearby: Mount Pozhyzhevska, 1,822 m · ≈ 390 m from the trail
05/ 06

Next.js

Shows it all in the browser: the map, the elevation profile and place cards — just as easy on a phone.

by the trail: Mount Breskul, 1,911 m
06/ 06

Hoverla · 2,061 m

The highest point of Ukraine. Total climb on the route: 898 m.

DEM profile: 1,352 → 2,040 m · official height 2,061 m
Hoverla 2,061Breskul 1,911Pozhyzhevska 1,822Dantsizh 1,850 Zaroslyak
Relief and the real route line, 215 points
1,400 m1,600 m1,800 m2,000 m
01PostGIS
02GraphHopper
03Elasticsearch
04Symfony
05Next.js
06Hoverla
01 km2 km3 km4 km5 km6.01 km
2,707routes
1,672,703GPS track points
660peaks
2,366stays

The profile shows real elevations from the digital elevation model (DEM) at the 215 points of the GraphHopper route; at the summit the DEM gives 2,040 m, the official height of Hoverla is 2,061 m. Figures from karpaty.rocks.

Technology already working in production

Not demos or prototypes: each one runs in a product people use every day.

01 · AI agent

An agent instead of search

People ask in plain language, and the agent calls the tools itself: water search in Elasticsearch, the bite forecast, the depth map and directions to the ramp. The server builds result cards from that data, so every answer is based on facts, not guesswork.

03 · Data graph

The Carpathian data graph

Every place is linked to others: a peak to its routes, a route to springs and shelters, a guesthouse to its village and the nearest trails. It powers “Nearby”, the trip planner and the schema.org markup that search engines and AI assistants read.

04 · Semantic search

Search that understands meaning

“A quiet place with a hot tub for two near the forest” finds the right guesthouses even when the description shares none of those words. Keyword matches and meaning (e5 embeddings) are combined into a single ranking — search on Rybalka Club works the same way.

05 · Speech analytics

Every call — scored, with evidence

Each recording is split by speaker, then an LLM scores the sales manager against your criteria and backs every score with an exact quote from the call. All calls are searchable by meaning (pgvector).

06 · Depth maps

Depths you can’t see from the shore

Depth contours are built from official USACE, NOAA and state soundings at adaptive intervals and never show water deeper than was actually measured. Where there is no survey, the map stays empty instead of guessing. GraphHopper plots the route from the ramp to the right depth.

07 · Tile server

Our own tile server

A dedicated Python + PostGIS server downloads the raw sources itself (NHD outlines and eHydro soundings), builds depth contours and packs them into vector tiles with tippecanoe (PMTiles). nginx serves them as plain static files via Range requests: full detail up close, lightweight tiles when zoomed out, and no CDN or object storage.

08 · E-commerce

E-commerce at scale

Edit a whole Shopify store like a spreadsheet: thousands of products, warehouses and price rules, every change previewed and every save undoable. AI assistants do the same over MCP. The product feed drives a Google Ads automation engine, where bid rules protect the budget and Claude writes the ad copy.

09 · Stablecoin payments

A payment that confirms itself

A freelancer issues an invoice in USDC, USDT or DAI, and the client pays over a low-fee L2 network. A service that watches the blockchain marks the invoice as paid the moment the funds arrive and emails a PDF receipt.

10 · Generative content

Content that writes and translates itself

AI parsers collect the raw data, an LLM writes the descriptions, a pipeline translates them into other languages, and image models create the illustrations. An editor only reviews the result — that’s how the images and copy on this site were made.

TripsQuick. Ask the water.

TripsQuick turns government soundings into one answer: where, how deep and when to fish today.

9:41
Lake Tahoe
Depths Bite
Morning bite
November 4, 2026, 6:40 AM
Tahoma
Area
123k ac
Max depth
1,640 ft
Ramps
5
Fishing assistant

Pressure is falling and the water is 11 °C, so bass will move closer to shore. Try the 8–12 m drop-off near Tahoe Keys, 4.4 miles from the ramp.

Water temperature
11.4°C
Bite score
74/100
12%
78,208
waters in the database
lakes, rivers and bays · USGS NHD
12,945
depth maps
from official soundings · USACE · NOAA
40,204
ramps & piers
water access · FWC · RIDB
19.6M
water-quality measurements
temperature, oxygen, pH · USGS · EPA

The hard part —
invisible to the user.

Nine public sources in different formats are merged into one page per lake or river. Where there is no official survey, depth is left blank rather than invented.

01 · Sources
USGS NHDNOAA chartsUS Army CorpsFWCGeorgia DNRWater Quality PortalRIDBOpenStreetMapWikimedia
02 · Core
Symfony · 30+ CLI pipelinesETL
Elasticsearch · kNNsearch
Depth interpolation from soundingsGIS
PostgreSQL · Redis · GraphHopperdata
03 · Product
  • Next.js · ISR · JSON-LD
  • Our own MapLibre vector tiles
  • AI assistant instead of search
  • Route to the ramp
  • Expo mobile app

Every product has a business model. Here’s how they make money.

I don’t just write code — I design how a product makes money and build that into the architecture from day one.

A trail along a Carpathian ridge leads to a summit cairn above a sea of clouds at dawn

Got a complex product?
I’ll take it to the summit.

End-to-end delivery, an architecture audit or a senior engineer for your team. Working remotely from the EU (CET).