AI · Every day

I build with AI every day.

Not hype. Not a demo reel — a tool I reach for every day. It writes code beside me, and it ships as real features inside the products I build. Here is exactly how I use it, where it fits, and the rules I keep it within.

Talk about your project See the projects

1Framea sharp prompt
2DraftAI proposes
3VerifyI review every line
4Shipto production
5Measurekeep or cut

My AI loop↳ AI does the typing — I own every decision

How I use it

Four ways AI earns its place

One tool, four very different jobs — from writing code beside me to features your users actually touch.

As a pair engineer

Claude Code sits in my terminal as a second senior engineer — it reads the whole repo, follows my conventions, and never tires of the boring parts.

What it does

Scaffolds features, refactors, and framework migrations
Writes tests and reproduces bugs from a stack trace
Reviews diffs and explains unfamiliar code
Automates repetitive work — codemods, config, glue

What I get

Days instead of weeks on repetitive work
More time on architecture and the hard calls

Claude CodeRefactorsTestsReviews

AI inside the product

AI features users actually touch — search that understands meaning, assistants that answer from your data, calls turned into searchable transcripts and scored.

What I build

RAG & semantic search over your content
LLM features: assistants, scoring, parsing, extraction
Speech-to-text with speaker separation
Grounding checks so answers stay backed by sources

What you get

Features that feel smart, not gimmicky
AI that does a job — measurable, not a demo

RAGEmbeddingspgvectorLLM

Content & media, faster

AI removes the blank-page problem. Copy in three languages, social images, and video — all produced within the same visual system, not generic AI output.

What it does

Multilingual copy & SEO (EN / ES / UK)
OG images and brand visuals generated to spec
Short videos and showreels from scripts
This very page — drafted, then hand-finished

What I get

A blank page is never the bottleneck
On-brand output, produced in hours

CopyOG imagesVideoi18n

Automation of the boring

Repetitive back-office work that used to eat hours — handled by small, reliable AI workflows that run themselves.

What I automate

Ad-copy generation and bid rules (Google Ads)
Lead routing to Telegram / WhatsApp / email
Internal tooling and one-off data jobs
A team AI Hub so wins get reused, not rebuilt

What you get

Hours back every week
Fewer manual, error-prone steps

AutomationGoogle AdsTooling

In production

AI I have actually shipped

Not slides. Not prototypes — real AI features running in production today. Explore any of them through the code or live product.

Speech + scoring

CallLens

Sales calls turned into coaching insights — speaker-diarized transcription, an LLM scorecard backed by quoted evidence, and every call semantically searchable.

Deepgram speech-to-text with speaker separation
LLM scores each rep against a configurable scorecard
pgvector embeddings for meaning-based search

SymfonyDeepgramOpenAIpgvectorRepo + live

Hybrid RAG

Polylog

A multi-agent chatbot SaaS with hybrid RAG (BM25 + dense + RRF) and a grounding verifier that blocks ungrounded answers.

bge-m3 embeddings on Elasticsearch
FastAPI backend, Next.js dashboard
A grounding verifier gates every answer

FastAPIRAGElasticsearchLive

Assistant instead of search

TripsQuick

An AI assistant that replaced site search — every answer grounded in our own depth surveys and public records, never the open internet.

e5 embeddings + kNN over measured facts
Result cards built server-side from Elasticsearch
Geocoding → driving route to the boat ramp

Next.jsSymfonyElasticsearchLive

AI assistant

Rybalka Club

A fishing community with an AI assistant that answers fishing questions, grounded in the handbook.

e5-large embeddings for semantic search + RAG
Elasticsearch as the only datastore
Next.js + Symfony

Next.jsEmbeddingsRAGLive

Guardrails

How I keep it honest

AI is powerful — and just as capable of being confidently wrong. These are the rules that keep it useful in production.

Human in the loop

AI writes the draft; I read every line before it ships. I own the decision, always.

Outcomes, not demos

If it doesn’t remove a real bottleneck or move a number, it doesn’t ship. Cool is not a reason.

Ground every answer

Answers stay tied to sources. A grounding check beats a confident hallucination every time.

The win is often deleting AI

Sometimes the best AI solution is no AI at all. A static page or a simple rule can beat a model. Killing a feature is a valid result.

Own the data, pin model versions

Self-hosted where it matters, exact model versions pinned — no silent drift in production.

Fast to test, ruthless to remove

Prototype in a day, measure honestly, keep only what earns its place.

The stack

What I reach for

The models, retrieval, and plumbing behind the work above — chosen for production, not demos.

Models & agents

ClaudeGPT (OpenAI)GeminiClaude Code

Speech

Deepgramdiarization

Retrieval

bge-m3e5-largepgvectorElasticsearch kNNRRF hybrid

Orchestration

Symfony MessengerFastAPISymfony WorkflowRedis queues

Let’s put AI to work.

An AI feature, RAG system, or workflow worth automating — tell me the problem, and I’ll tell you whether AI is actually the right tool.

vitaly@kindrakevich.com See the projects