Draft reply to a customer
AI reads the conversation and suggests a reply in the seller's own style. The seller edits it and sends it.
running in Kelisim
Roman Pogrebnyak
Integration engineer · Almaty · open to projects
I build websites and software, bring in AI and connect hardware for businesses
I write websites, CRMs and business systems in Laravel, Go and Python. I connect cameras, access control, barriers, scales and telephony to them, and add AI where it saves people time. I take a system all the way to real use, not just to a demo.
- websites and CRM
- ANPR cameras
- WIM scales
- lidars
- barriers
- bill acceptors
- Access control
- Asterisk / SIP
- Tuya locks
- public procurement
- Laravel · 1C · ESBD
- AI in WhatsApp
- AI bots
since 2013building systems and integrations
Laravel · Go · Pythonfrom driver to web panel
kk · ruinterfaces, voice and receipts
data from hardware and messengers in one system
AI for business
AI inside your processes
The assistant lives inside the tools people already use: WhatsApp, Telegram, the phone, your accounting system. Nobody has to learn a new app. AI prepares, a person decides.
Voice message transcription
Customers' voice messages become text you can search and reply to.
running in KelisimCustomer card
Deal stage, promises, risks and who needs a reply today. Built from the conversation automatically.
running in KelisimBookings and FAQs
A messenger bot answers customers around the clock and books appointments.
AI receptionistVoice in Kazakh and Russian
Text-to-speech for call-outs and alerts on your own server, no cloud.
emergency alertsTenders and supplier checks
Finding suitable tenders and checking counterparties against open data.
tender assistant- A person decidesAI prepares a draft or a hint, an employee sends it.
- Your data stays under controlPersonal chats are never read, and some tasks run on your own server.
- The right model for the jobClaude, OpenAI, Whisper: an inexpensive model for the daily flow, a stronger one for hard cases.
AI for sales · Kelisimused by real sellers
AI assistant for sellers in WhatsApp
- A customer writes or sends a voice note
- Kelisim reads only chats in the “Customers” list
- AI writes a draft in the seller's style
- The seller edits and sends it
Project site: kelicim.kz ↗
Sellers keep using their own WhatsApp. The assistant transcribes voice notes, suggests replies, keeps a customer card and reminds about promises. It never reads personal chats and never sends anything itself, so the risk of the number being banned is minimal.
How it's built · data flows top to bottom
- Channel
seller's WhatsAppWAHA · GOWSthe seller's own number as a linked device, no separate bot - Safety
waha_readonlyread-only and only chats in the “Customers” list: no sending, no read receipts - AI
Claude HaikuClaude OpusWhisperdrafts in the seller's style, “make it better”, voice transcription, a customer card with promises and risks - Server
appbridgePINa dashboard for each seller on kelicim.kz, a bridge to WhatsApp, PIN login
PythonJSClaude API · Whisper · Docker · Linux
What else Kelisim does
KelisimCard
Customer: beauty salonYour move
newtalksquotecontractpaid
Promises
✓ Send the price list
○ Quote by Friday
Risks
Comparing with a competitor
Next step: send the quote today by 6 pm
01 / 12
Customer card
Deal stage, promises, risks and whose move it is.
Example with sample data
KelisimToday
Your moveCustomer: clinicwaiting 2 h for a reply
Your moveCustomer: caféasked about delivery
Promise dueCustomer: beauty salonquote by 6 pm
Silent 3 daysCustomer: shoptime to follow up
Open the first
02 / 12
Today
Who needs a reply, which promises are due and who has been silent for 2+ days.
Example with sample data
KelisimClose it
Your offer10% off if paid by Friday
New termsDeadline“No” question
Option 1We hold the 10% discount until Friday. Shall we sign today?why: a clear deadline and a small step
Option 2If now is not the time, just say “no” and I won't rush you.why: removes pressure, so the customer replies
+ 1 more option
03 / 12
“Close it”
Three message options to move a customer to a decision, with why each one works.
Example with sample data
KelisimMoney
What this customer will bring≈ 120,000 ₸ a monthexample
Deal odds
55%
+10% replies quickly
+5% asked about the contract
−5% comparing prices
Stage “quote”: 40% base plus factors
04 / 12
Revenue forecast
How much a customer will bring and the deal odds by stage.
Example with sample data
KelisimVoice notes
▶0:42
“We need 10 pieces for the holiday, but delivery must be by Wednesday”
Voice notes summary
Wants: 10 pieces for the holiday
Worried about: delivery time
Agreed: quote tomorrow
05 / 12
Voice notes
Transcribed on our own server, plus a summary: what the customer wants and what was agreed.
Example with sample data
KelisimDocuments
PDFpricelist.pdfJPGshopfront photoXLSorder.xlsx
AI has read
Order: 3 items, 245,000 ₸
Photo: shopfront 2×1 m
Used in the draft reply
06 / 12
Photos and documents
AI reads photos, PDFs, Word and Excel files from the chat and uses them in replies.
Example with sample data
KelisimPrice list
Homepage banner25 000 ₸Mailing15 000 ₸Top of catalogue10 000 ₸
A homepage banner costs 25 000 ₸ per week.
✓ The price comes from the list, not from the AI
07 / 12
Price list and terms
Prices in replies come from your price list; AI never makes them up.
Example with sample data
KelisimHand over
To
Colleague 1Colleague 2
Dossierstage: quote · promise: quote by Friday · last 10 messages
Hello! I'm handing you over to my colleague, who will pick up right where we left off.
✓ Dossier sent to the colleague
08 / 12
Hand over to a colleague
A customer dossier and a ready “handing you over to my colleague” message.
Example with sample data
KelisimCalendar
Promise: call on Thursday at 11:00
Add to calendar
MonTueWedThu11:00 callFri
✓ Reminder 15 minutes before
09 / 12
Calendar
Meetings and promises go into the phone calendar with one tap.
Example with sample data
KelisimTeam
Seller 114 customers · 3 waiting
Seller 29 customers · 1 waiting
Seller 36 customers · all answered
You can see where customers get lost
10 / 12
Manager dashboard
A summary of every seller on the team on one screen.
Example with sample data
KelisimTip
Too expensive. Others are cheaper.
Price objectionAsk what the customer is comparing with and what matters more to them than price.SPIN method
I understand. Which offer are you comparing with? I'll check how we differ.
11 / 12
Sales coaching
Negotiation tips based on SPIN, Sandler and other methods.
Example with sample data
KelisimPause
on holiday until 20.11Customer: caféwrote: “back after my holiday”
Reminders
Won't show up in “Silent” or “Your move”
✓ Back in the list on 21.11 automatically
12 / 12
Pause
Customer on holiday or ill: reminders leave them alone.
Example with sample data
AI bots
AI receptionist for doctor appointments
A Telegram and WhatsApp bot understands patients' plain messages, matches them to a doctor by specialty and books the visit. It works nights and weekends, so front-desk staff can focus on patients in the clinic.
How it's built
- Channels
TelegramWhatsApp webhookpatient messages - Logic
message_handlerprompt_builderllm_responsereminder_schedulerunderstands the request, picks a doctor, sends reminders - Data
databaseclinic_infoPostgreSQL: doctors, chat history
PythonOpenAI · PostgreSQL · Docker
Public procurement · bigtender.kz
Tender assistant
Collects data from the public procurement portal and tax sources, keeps a database of contracts and companies, and rates supplier risk. Helps you find a suitable tender faster and check a counterparty.
How it's built
- Collection
scrape_goszakupscrape_contractsscrape_kgdows_clientaisoip_clientprocurement portal, contracts, tax data - Analysis
risk_insightsorganization_storecontract_storecompany and contract database, supplier risk - Interface
adminweb dashboard with login
Python
Hardware and integrations
How it works on site
Each animation shows how the system works for people. Below it, “How it's built” lists the system layers, the modules from the code and the languages I wrote it in. Plates, amounts and texts in the animations are examples.
Toll roads · KazTollrunning on site
Toll payment terminal in the lane
- The camera reads the plate
- Scales and lidar measure the truck in motion
- The terminal takes payment and prints a receipt
- The barrier opens
I wrote the hardware drivers and the lane service in Go on Linux: bill acceptor, coin board, induction loop, receipt printer in Kazakh and Russian, barrier, display board. The terminal checks itself every 15 seconds and takes itself out of service on a fault. Together with the WIM station (next project) this makes a complete truck lane.
How it's built · data flows top to bottom
- Hardware
NV200PRS17AGEA2488VKP-80IILBA 63PGdisplaybills, coins, loop, receipt printer, barrier - Drivers
serialportcashprs17ageaiopanelprintermodbusbarrierledsignRS-232, Modbus TCP, ESC/POS; receipts print as images so Kazakh letters render - Logic
lanehealthjournalpassage, change in coins, self-check every 15 s, cash log - Link
apiD-BusHTTP API and message bus for the road operator's system - Install
discoverconfigdeployauto-detects devices on ports, one-command install, systemd with watchdog
GobashRocky Linux 9 · systemd · SELinux · Docker for builds
Toll roads · WIM
Weigh-in-motion station for trucks
- The camera reads the plate and takes a photo
- Sensors in the road weigh every axle
- The lidar measures size and counts axles
- The service merges it all into one vehicle record
The truck drives through without stopping, and the station collects everything three devices from different vendors can tell about it. Each device answers with its own delay: the camera sees the plate early, the lidar answers once the truck has fully passed. The service matches events by lane and time window, corrected for each device. If the scales count 3 axles and the lidar 2, the record is flagged for review instead of being lost.
How it's built · data flows top to bottom
- Devices
Kistler LineasIntercompVanJee WLR-711Tattile Vega53weight sensors in the road, a lidar beside the lane, a plate camera - Intake
tattileintercompvanjeethe camera pushes JSON over HTTP, the scales use TCP (WIMLOGIX), the lidar an SDK; raw data is stored in full - Matching
matchthree streams into one vehicle by lane and time window - Record
passagelimitsaxle weights, dimensions, plate, photos, mismatch flags; weight and size limit checks - Service
wimdHTTP receivers, health checks, record storage
GoLinux · HTTP · TCP · JSON
Alerts · Emergency servicescompleted project
Voice call-out to staff with confirmation
- The operator types the alert
- The system voices it in Kazakh and Russian
- Calls go out: staff press 1
- Up to 3 retries and a list of who missed it
Own telephony on Asterisk via the carrier's SIP trunk. Speech synthesis runs on our own server, so staff data never goes to the cloud. By estimate, 1,400 people are called in about 15 minutes on 100 lines.
How it's built · data flows top to bottom
- Panel
staticemployees.csvalert text, staff list with language: kk, ru or both - Backend
app/api/campaigns/api/employeescall campaign, call queue, retries, report - Voice
ttsPiperKazakh and Russian speech synthesis on our own server - Telephony
extensions.confPJSIPfunc_curlcall via the carrier trunk, key press 1, events back to the backend
PythonAsterisk dialplanFastAPI · SIP · Ubuntu
Education · Safe Schoolcompleted project
Card entry, cameras and parent notifications
- A student taps a card
- The turnstile opens, the camera takes a photo
- The event goes to the school system
- The parent sees the child has arrived
One system for entry, video, the panic button, a canteen paid from the student's balance, and parent notifications. Access terminals from three vendors and IP cameras are connected. The interface is in Kazakh, Russian and English.
How it's built · data flows top to bottom
- Hardware
HikvisionDahuaSigurIP camerasaccess terminals from three vendors, video cameras - Services
HikvisionAccessControlDahuaAccessControlSigurServiceCameraServicea separate adapter for each vendor - Video
rtps-hlsFFmpegServiceWebRTCServiceRTSP from cameras to the browser, a separate Go module - Features
EmergencyAlertMealNotificationAnnouncementpanic button, canteen, notifications, announcements - Admin
FilamentJobsEventsadmin panel, background queues and events
PHPLaravelGoMySQL · Redis · FFmpeg · Docker
Customs · asid.kzcompleted project
Inspectors' body cameras and a monitoring centre
- The camera records at the post
- A live stream or clips go to the server
- The centre watches in real time
- Video is linked to the declaration in ASYCUDA
Video from body cameras arrives two ways: a live stream or uploaded recorded clips. Video storage, user roles, an event log and a link to the ASYCUDA customs system.
How it's built · data flows top to bottom
- Camera
streamclipslive stream or uploading recorded clips after a shift - API
authvideosadminlogin and roles, video in and out, managing cameras and people - Storage
PostgreSQLMinIORedisalembicmetadata and events in the database, video files in object storage - Integration
asycudalinking video to the declaration in the customs system
PythonFastAPI · Docker
Smart locks · Tuyacompleted project
Smart lock access log in your own system
- The lock is opened by code, card or app
- The event goes to the Tuya cloud
- Our service fetches it via the API
- An entry in your system's log
Lock status and opening history for any period in your own CRM, not just the vendor's app. Bookings, temporary passcodes, Telegram notifications and accounting. A fit for short-term rentals, offices and residential complexes.
How it's built · data flows top to bottom
- Lock
Tuyacode, card or app - Cloud
tuya_clienttuya_cachelock_cryptoTuya OpenAPI, AES encryption per the lock protocol - Services
lock_servicebooking_servicepassword_storelock control, bookings, temporary passcodes - CRM
crmtelegram_serviceaccounting_serviceadmin, notifications and accounting
PythonFastAPI · httpx · systemd
Laravel development
Business systems, CRMs and platforms
My main tool for web systems since 2016: from document management and insurance products to billing and marketplaces. I write the backend, the admin panel and integrations, and deploy with Docker.
Insurance · ESBD
Online insurance integrated with ESBD and 1C
Online policy pricing and sales, exchange with the Unified Insurance Database (ESBD) over SOAP and with the company's 1C. Built for three insurers, including 8 online products from scratch at one of them.
Laravel · SOAP · 1C · RabbitMQ · Kafka · Go
B2B + B2C · sauda.kz
High-load trading platform
A marketplace for wholesale and retail. Fast catalogue search, background queues, real-time updates, monitoring and error tracking.
Laravel 11 · Octane · PostgreSQL · Redis · Meilisearch · Reverb
Utilities
Utility billing
Water and heating charges for property managers and homeowner associations. Buildings, meters, five calculation methods, penalties, Excel reports, a customer portal.
Laravel 11 · Filament 3
Service
Help-desk CRM
Tickets are assigned by engineer workload. Time tracking, hourly billing reports, roles and permissions.
Laravel 13 · React 18 · TypeScript
Sales
“KRAFT Proschet”: from request to register
The deal chain: request, rates and costs, commercial offer, supplier order, register. Only the director approves offers.
React · Python · FastAPI · PostgreSQL
Building materials
Material Hub: B2B finishing-materials catalogue
A filterable catalogue of tiles, stone, lighting and decor. Automatic import of catalogues from suppliers present in Kazakhstan.
Laravel 13 · Inertia · React · MySQL
Also: payment integrations (TipTop Pay), car loans, car and motorbike marketplaces, a CRM with Asterisk and call history, moving a monolith to microservices.
How I can help
What I can do for your business
AI assistant for sales and support
- Draft replies to customers in WhatsApp and Telegram
- Voice transcription, customer cards, reminders
- A receptionist bot for bookings and FAQs
for: sales teams, clinics, service companies
Connect hardware to software
- Cameras, access control, barriers, scales, lidars, cash devices, locks
- Drivers, protocols, one-command install
- Self-checks and clear errors for field technicians
for: roads, parking, warehouses, schools, residential complexes
Automation with local systems
- Public procurement and tax data
- Telephony and SIP with Kazakh carriers
- Dashboards and reports for managers
for: government bodies, enterprises, suppliers
A Laravel web system built around your process
- CRM, accounting, billing, customer portals
- Integrations with 1C, ESBD and payment systems over REST and SOAP
- Improving and migrating legacy systems
for: insurance, service and trading companies
- ScopingA call or meeting: what the site or process is and what gets in the way today.
- PilotOne lane, one seller or one entrance. Results on your own task.
- LaunchOn-site install, staff training, a guide for technicians.
- SupportI monitor the system and extend it as you grow.
Experience
10+ years in development and integration
I started with websites, then spent years connecting insurance and accounting systems to government databases, 1C and telephony. My recent projects are about hardware and AI.
- Go
- PHP · Laravel
- Python · FastAPI
- Vue · React
- PostgreSQL · MySQL
- RabbitMQ · Kafka
- Docker · CI/CD
- SOAP · REST
- 1C
- Asterisk
Education: Kazakh-Russian International University, Aktobe, 2009. Information technology.
- nowOwn projects and client work
Toll payment lanes and WIM for toll roads, the Kelisim AI assistant, alerts, schools, locks, tenders.
- 2025Craft Promotion · remote
Full stack: web services in Laravel and Vue, Docker, RabbitMQ.
- 2022 – 2024Chocolife.me · senior PHP developer
Moving a monolith to microservices, external service integrations, CI/CD. Wrote a Go bot that reports build statuses.
- 2019 – 2021Insurers: Eurasia, Sinoasia B&R, Oil Insurance Company
Online insurance and integration with ESBD (SOAP) and 1C. An auth microservice in Go. A CRM for managers with Asterisk and call history.
- 2016 – 2018Kamal Oil · Uralsk
Moved a document management system to Laravel, linked PDMS with a web portal.
- 2013 – 2015Web-Design · Uralsk
Websites, SharePoint, iOS apps.
Contacts
Tell me what you want to solve
I answer personally. Keep it short: what the site or process is and what gets in the way.