Systems designed, built and deployed end to end. The metrics below come from client analytics.
Selected work
Four systems, four industries, numbers from the clients.
Every project below was designed, built and deployed end to end. Metrics come from client analytics.
E-commerce · Auto parts
AI sales agent over a catalogue of thousands of SKUs
Customers asked about stock and fitment for a specific car model. Managers could not keep up and replied in 2-3 days. Nights and weekends went unanswered, so inquiries went to competitors.
An AI agent with RAG search across the catalogue answers stock, price and compatibility instantly, clarifies the configuration and suggests an original part or an equivalent. Trigger rules route the conversation to a human and create the deal in Bitrix24.
AI Agent
RAG
Telegram
Website
Bitrix24
2-3 days → 10 min
Response time
+75%
Conversion to order
-60%
Manager workload
Education
Enrolment agent for a course launch
The agent closes standard questions, guides the student through the funnel to payment and hands over only the cases that need a human.
24/7
Student response
up to 70%
Less routine work
+10-25%
Conversion
Expert services
First-touch qualification across three channels
The agent takes first contact in every channel, qualifies the request and passes on only warm, pre-briefed clients.
24/7
Inquiry handling
-65%
Less routine work
+22%
Warm leads
Service center
Intake and repair-status assistant
The agent accepts requests, answers on timelines and status, and removes the repetitive load from the reception team.
A sample of what I have designed, built and handed over. Client names are withheld, the engineering is described as built.
Retrieval agent
Sales agent over a catalogue of thousands of SKUs. Vector search answers stock, price and fitment for a specific car model, then writes the deal into Bitrix24.
CatalogueVector searchAgentCRM
Qdrant
RAG
Bitrix24
Omnichannel desk
Telegram, VK and WhatsApp landing in one inbox. The agent replies, the operator takes over mid-conversation without losing context.
TG · WA · VKSingle inboxOperator
WhatsApp API
VK
Telegram
Payment safety net
A guard layer between the payment provider and the funnel: idempotent processing, amount reconciliation, retries with backoff, and an alert when an order hangs unpaid.
PaymentReconcile · retryFunnelAlert
Python
Webhooks
Alerting
Knowledge base
A dealer support base built from a source PDF: 80 questions across 12 sections. The agent answers from the document only and says so when the answer is not there.
PDFChunksEmbeddingsAnswer from doc
Embeddings
Vector search
Meeting intelligence
Hundred-minute calls turned into structured reports in Telegram. Long audio is chunked and reduced so it fits provider rate limits without failing.
AudioChunksMap-reduceReport
Speech to text
Map-reduce
Marketplace operations
Seventeen scenarios keeping prices and product content in sync on a marketplace, including discount bases and quarantine handling.
Price · content17 scenariosStorefront
Marketplace API
n8n
Content pipeline
Idea to script to approval to publication, with a human approving in one tap from a chat message.
IdeaScriptApprovalPublish
LLM
Approval flow
Lead engine
A niche landing page, a blog and a quiz all write into one CRM, with deduplication by phone and messenger id so the same person never becomes two leads.
Landing · blog · quizDedup · scoringCRM
PostgreSQL
Deduplication
Why speed converts
The first minutes decide who gets the deal.
The money leaks between the inquiry arriving and a human seeing it. Three published findings on what that gap costs:
×21
Higher odds of qualifying a lead when you answer within 5 minutes instead of 30.
MIT / InsideSales, 2007
78%
Of deals go to the company that responds first.
Harvard Business Review, 2011
-80%
Drop in lead quality after just 5 minutes of silence.
Lead Response Management
The five-minute window
Answer in five minutes and you are twenty-one times more likely to qualify the lead.
The same inquiry, the same team, the only difference is how long it waited. Published odds of reaching and qualifying a lead:
Replied within 5 minutes×21
Replied after 30 minutes×1
MIT / InsideSales lead response study, 2007
Odds of qualifying a lead by response delay
Replied within 5 minutes
×21 × odds of qualifying
Replied after 30 minutes
×1 × odds of qualifying
Currently accepting new projects
Which part of your day would you hand to a system first?
Tell us what eats the most time right now. On the call you get a map of what can be automated, in what order, and what it costs. If the honest answer is that you do not need this yet, you will hear that too.