01Before · Implemented · After
Three implementations inside a live business. All three still in daily use.
Each one started the same way: find the manual workflow that costs the most, quantify it, then decide whether to automate it or redesign it. A company running ~150 deals a month gets no maintenance window — every change had to work while the business kept selling.
A1
Multi-channel lead intake — automated capture into CRM
2025–2026
Before
Inbound leads landed across several disconnected channels and were transcribed into amoCRM by hand. Leads were missed, data-entry errors reached the pipeline, and three reps lost a third of every morning to triage instead of selling.
→
Implemented
- Mapped the intake workflow end to end and quantified the cost of each manual step
- Built source parsing and automatic deal creation in amoCRM, with validation and exception handling on the fields sales needs at quoting time
- Redesigned task rules and Salesbot automations around the process as it actually runs
→
After
≈3 h/day
returned to each sales manager, by my operational estimate. The team redirected that time into client work, with fewer missed leads and manual-entry errors.
Adoption: full sales team migrated and still runs on it. I wrote the SOP and trained the team.
amoCRMsource parsing
workflow automationSalesbot
A2
Commercial reporting — AI agents over live business data
2025–2026
Before
Sales, supply, inventory turnover and revenue analysis ran off manual marketplace exports and consumed a week or more of the owner's time every cycle. Decisions waited on the report.
→
Implemented
- Deployed AI agents (Claude Code, Codex, Hermes) against company databases and marketplace report exports
- Defined the decision questions the business actually asks, so the output drives action instead of producing another dashboard
- Built the runbook and trained the company owner to operate it himself
→
After
−70%
faster reporting cycle — from over a week to two days. Commercial decisions now move at the speed of the business.
Adoption: the owner runs it unaided. Delivering a tool only its author can operate is not an implementation — it is a dependency.
AI agentsClaude CodeCodex
Hermesstakeholder training
A3
Marketplace content — vendor replaced by an in-house AI pipeline
2026
Before
Every product card came from an external design vendor: brief, wait, invoice, repeat. Card performance was never measured, so spend continued with no evidence of what actually sold.
→
Implemented
- Ran A/B tests across the existing catalogue to identify which designs converted
- Built a training set from the top-converting cards — selection driven by commercial data, not preference
- Trained and deployed a Style LoRA reproducing that proven style locally, and handed the tool to the marketplace manager
→
After
Vendor
dependency removed. The marketplace manager now produces and ships cards directly, at the style the data proved converts.
Commercial logic: the training images were selected from A/B-tested product cards. The goal was a repeatable content workflow the marketplace manager could own.
A/B testingdataset curation
Style LoRAComfyUIIdeogram 4
Client delivery and team-built products.
A clinical practice, a service centre and a community-events product. Different domains, same approach: understand the routine, build the useful part, and hand it over.
C1
A clinician's method, running as a live screening tool
2026
Before
Parents arrived at a first appointment with a vague description — "he won't listen, school is going badly". The specialist spent the session extracting a structured picture instead of working with it, and appointment time is a practice's most expensive resource.
→
Implemented
- Encoded the specialist's existing framework — Luria's three functional blocks — rather than inventing my own scoring: 17 questions on the energy block, 10 on information processing, 15 on programming and control
- Designed around the real user, a tired parent on a phone: nine school-related questions are hidden for pre-school children, and a back button removes the previous answer instead of just scrolling
- Different thresholds per age band, and a tie-break order between blocks so an ambiguous result resolves by clinical weight rather than by whichever block came first
→
After
42
questions, scored per block. Approximately 60+ families have used it, according to the client; the specialist receives a profile before consultation.
Live and taking clients. Two outcomes the specialist reports: less appointment time spent gathering history, and parents who now reach the consultation on their own. The tool structures a parent's observations for a qualified professional — it does not diagnose, and is not presented as doing so.
domain modellingNode.js
Telegrafsurvey UX
C2
Customer needs captured before a manager picks up
2025–2026
Before
Every enquiry for a custom PC build started with the same round of questions — budget, use cases, component preferences, deadlines. Managers repeated it by hand with every customer before they could propose anything.
→
Implemented
- Mapped the qualification conversation managers were already having, then moved it ahead of first contact
- The bot collects budget, usage patterns, component preferences and timing, and delivers a customer profile into the team's working chat — the main need identified, not raw answers
- Deliberately stopped short of recommending builds: the routine moves to the system, the expertise stays with the specialist
→
After
170+
customers qualified in six months, approximately, according to the client. Managers receive a structured brief before first contact.
Delivered and closed. This is the shape most automation requests actually take: not replacing the expert, but removing the identical opening conversation they have three hundred times a year.
lead qualificationrequirements discovery
Telegramhandover to sales
C3
Community events — from a poll to a measurable attendance funnel
2026
Before
Across 11 community meetups, people pressed "Yes" 45 times — and nobody ever counted how many arrived. 148 answered "Maybe", more than the yes votes, and that segment was never worked at all. The organiser could not tell how many to cater for, or whether anything was improving.
→
Implemented
- Product Owner in a team of three on weekly sprints with daily stand-ups — Jira, Slack, GitHub, Google Workspace
- Designed the funnel Maybe → confirmed → attended → rated → returned, measuring confirmation and attendance as separate events
- "I'm here" appears one hour before the meetup and not earlier, so it measures attendance rather than intention; only attendees can rate, so scores are not diluted by people who saw an announcement
→
After
1 day
from zero to a working MVP on a permanent public address — with a way to measure turnout that the organizer did not have before.
Partnership with a stake in the product. The hard part was not the build — it was refusing to measure intention when everyone else does. Attendance is the number the organiser actually needs.
Product OwnerJira · Slack
weekly sprintsFastAPI · React
B1
The operating system underneath — eight years of commercial systems
2018–2026
Before
Joined as the company's only sales manager. Sales, procurement, warehouse and logistics coordinated by memory and messages, so reps quoted dates the warehouse could not meet.
→
Implemented
- Rebuilt the lead-to-order workflow in amoCRM — pipelines, stages, task rules, telephony and messenger integrations
- Integrated amoCRM with MoySklad ERP over REST APIs: live per-SKU stock at quoting time, a picking queue that stopped losing assembly requests, one shipment view for logistics
- Launched three national marketplace channels from zero — Ozon, Wildberries, Yandex Market — growing to ~35% of company revenue
- Scaled the team from 1 to 8 and built the enablement layer: sales scripts by client type, onboarding, SOPs, pricing rules and the KPI scheme. New reps productive in 2–3 weeks
→
After
5.4×
company revenue growth over eight years — RUB 21M to RUB 114M — a company and team result across B2B, marketplace and distributor channels.
Why this is the foundation: knowing where a commercial operation actually breaks is what makes an AI implementation land. I spent eight years inside the process before automating it.
amoCRM × MoySklad ERPAPI integration
Ozon · WB · Yandex Marketprocurementteam 1→8