Максим

MAKSIM OPALINSKYI
Python Backend Developer
Vinnytsia, Ukraine | [открыть контакты](см. выше в блоке «контактная информация») | [открыть контакты](см. выше в блоке «контактная информация») | github.com/hqrcules

SKILLS PROFILE
Languages Backend engineer building production Python systems that handle real money, real
Python · SQL · JavaScript users, and real-time traffic. Shipped a 7-partner ticket aggregator processing live
Backend Frameworks Monobank payments, an OpenAI Realtime tutoring backend with sub-400 ms voice
FastAPI · Django · Django REST latency, and an internal HR assistant serving 350+ employees. Comfortable in both
Framework · Django Channels · Pydantic FastAPI / async and Django / ASGI stacks; my work usually starts at “integrate this
Async & Concurrency external API” and ends at “survive production traffic on it.”
asyncio · httpx · ThreadPoolExecutor ·
ASGI (Daphne, uvicorn) · APScheduler · SELECTED PROJECTS
distributed locking
Databases & Caching nashaplatforma — Bus Ticket Aggregation Platform DEC 2025 — PRESENT
PostgreSQL · MySQL · SQLAlchemy Stack: Python, FastAPI, Pydantic, asyncio, httpx, ThreadPoolExecutor, Redis, PostgreSQL,
(async) · Redis (cache, Pub/Sub, locks) · aiogram, Monobank API, XML / JSON / REST, Docker, Railway
Whoosh • Architected and integrated 7 third-party bus ticket APIs (LikeBus, SVS, BlaBlaCar,
AI / NLP EWE, Gunsel, Pavluks, NashaPlatforma) with heterogeneous protocols (XML,
OpenAI API · OpenAI Realtime API · JSON, REST) into a single unified API layer, collapsing frontend and bot
Langfuse · regex intent routing · integration to one endpoint per action.
transliteration & morphology • Built concurrent route search across all partners simultaneously using
Audio & Parsing ThreadPoolExecutor, delivering aggregated results from 7 sources in under 3
WebRTC VAD · FFmpeg · pydub · lxml · seconds.
BeautifulSoup · XML / JSON / REST / SOAP • Implemented a two-layer Redis caching strategy (view-level + per-route) with
Integrations distributed locking to prevent thundering-herd on popular routes, cutting
Telegram Bot API (aiogram, Pyrogram, redundant partner API calls by up to 90%.
Telethon) · Monobank · MS Bot • Designed full order lifecycle across 20+ REST endpoints: route search, seat
Framework · Adaptive Cards · OAuth · selection, order creation, Monobank payment confirmation, ticket delivery,
JWT · Webhooks cancellation and refund handling — with idempotency checks preventing double
Real-time charges.
WebSockets · Django Channels · Redis • Built an advanced passenger pricing engine supporting 6 passenger types
Pub/Sub · FSM-driven flows (adults, children, seniors, disabled, animals, students) with per-partner discount
DevOps & Tooling logic that automatically filters out incompatible routes.
Docker · Docker Compose · Nginx · • Shipped a parallel customer channel as a multi-step Telegram bot in aiogram
GitHub Actions · Ansible · Railway · with FSM-driven flow across 9+ screens, persisting state in Redis and reusing the
DigitalOcean · PM2 · Postman · Git same partner integration layer as the website — zero code duplication between
channels; external carriers route through backend order_new → Monobank →
EDUCATION order_confirm, while NashaPlatforma is handed off to website checkout via
deep-link with prefilled passenger data.
Vinnytsia National Technical University
• Eliminated a legacy double-spend bug where the bot was triggering partner
Bachelor of Computer Science
2023 — 2027
ticket purchases pre-payment immediately after seat reservation, burning agent
funds on unpaid bookings — consolidated confirmation to a single post-payment
JavaRush — Python University converging point on the backend.
Intensive Python backend program • Built a dedicated notification subsystem on a separate Telegram bot,
2024 — 2025 broadcasting confirmed orders with HTML-formatted route / passenger /
payment details and best-effort partner ticket-ID extraction across 7 response
LANGUAGES formats — hardened with per-user error isolation, graceful 403 handling, and
Ukrainian — Fluent (native) stage-aware token loading to keep prod and dev channels strictly separated.
English — Intermediate • Maintained per-partner dev/prod configuration with stage-aware cache keys
and base URLs (test-api vs prod-api), allowing the same binaries to serve both
environments without code changes; deployed on Railway with an automated
redeploy pipeline.
AI Lingua — Language-learning platform with realtime voice AI MAR 2026 — APR 2026
Stack: Python, Django 5, Django Channels, Daphne (ASGI), Redis, PostgreSQL, OpenAI Realtime
API, WebRTC VAD, FFmpeg, pydub, drf-spectacular, Langfuse, Docker Compose, Ansible, Nginx
• Designed and built a Django 5 backend integrating OpenAI Realtime API over
WebSockets to deliver sub-second voice-to-voice tutoring sessions.
• Implemented a bidirectional audio streaming pipeline on Django Channels +
Daphne (ASGI) with Redis as the channel layer, sustaining concurrent realtime
dialogs with under 400 ms turn-around latency.
• Engineered a custom voice activity detection layer with webrtcvad and pydub,
trimming silence before forwarding audio frames upstream and reducing wasted
OpenAI tokens by ~30%.
• Developed a TTS prefetch service and dialog orchestrator that pre-generates the
next AI response while the user is still speaking, cutting perceived response
delay by 45%.
• Built a demo-mode subsystem allowing anonymous users to try lessons without
registration, persisting progress in PostgreSQL JSONFields and migrating it into
real accounts automatically on signup via Django signals.
• Authored 40+ REST endpoints documented with drf-spectacular (OpenAPI 3)
covering lessons, vocabulary, dialog scripts, billing and admin flows.
• Integrated Langfuse for end-to-end LLM observability — cost-per-session
tracking and prompt-version regression checks across all AI features.
• Implemented multi-language content with django-modeltranslation, social auth
via django-allauth (Google, Apple), and FFmpeg-based audio post-processing for
vocabulary-extraction-from-video features (YouTube + uploaded clips →
structured learning material).
• Containerized the stack (Django + Daphne + PostgreSQL + Redis + Nginx) with
Docker Compose and automated staging / production provisioning with Ansible
playbooks.

Real-time Telegram CRM Ingestion Backend — FastAPI bridge to a NestJS
WebSocket layer JAN 2026 — FEB 2026
Stack: Python, FastAPI, Pyrogram, Telethon, SQLAlchemy (async), PostgreSQL, Redis Pub/Sub,
NestJS (consumer), DigitalOcean, PM2
• Architected a real-time Telegram CRM ingestion backend on FastAPI +
Pyrogram/Telethon, bridging multiple live Telegram user accounts into a NestJS
WebSocket layer through Redis Pub/Sub — operators see incoming messages
and lead state changes in the UI immediately, no manual refresh.
• Designed a six-channel Pub/Sub event bus (chats:new/update,
messages:new/update, leads:new/update) consumed by two NestJS gateways,
decoupling Python ingestion from Node delivery so each side scales
independently and keeping PostgreSQL as the single source of truth.
• Built an auto-lead pipeline that materializes a CRM lead on the first inbound
message from a new contact, hydrates it with the owning manager and fully
resolved status object, then routes leads:new to a per-manager room (lead:user:
{managerId}) — removing manual lead creation for inbound traffic entirely.
• Diagnosed and eliminated a first-message race condition where the frontend
was dropping message:new events arriving before the chat existed in the UI —
fixed by emitting chat:new as a broadcast first and gating the follow-up
message:new behind a short delay, applying the same ordering across cross-
account chat discovery.
• Engineered the lead status transition flow end-to-end (Новий → В роботі on
first manager reply) with a single DB write and a leads:update payload carrying
the complete status object — fixing a production bug where NestJS was silently
discarding status updates because status was arriving as null.
• Hardened SQLAlchemy async transactions against greenlet_spawn errors caused
by accessing expired attributes after commit, by restructuring publishers to
snapshot every field the Redis payload needs before the commit boundary.
 • Implemented a multi-account Telegram manager wrapping Pyrogram clients per
Nest-side account with lazy session loading, idempotent handler registration to
prevent duplicate listeners on reload, and a single FastAPI lifespan-managed
startup/shutdown sequence; deployed on a DigitalOcean VPS under PM2.

Enterprise HR Operations Assistant — Internal AI chatbot over enterprise HRIS APR
2026 — PRESENT
Stack: Python, FastAPI, Whoosh, Microsoft Bot Framework, Copilot Studio, Adaptive Cards,
APScheduler, HRIS REST API, regex-based NLP routing
• Shipped end-to-end an intent-driven NLP routing engine handling 15+ query
types in a single FastAPI endpoint (profile fields, salary, vacation balance, active
leave, absences, birthdays, tasks, requests, KB search, category listings,
organizational lookups) — used in production by 350+ employees across 20+ KB
categories.
• Built a production-grade Ukrainian / English bilingual search engine on Whoosh
with custom analyzers: bidirectional Cyrillic↔Latin transliteration, 4-char prefix-
stem matching for morphological declensions, brand and cross-script synonym
groups, title field-boost, and topic-aware overview-article boost gating — raising
top-1 accuracy from ~50% to 95%+ across audited query suites.
• Designed a whitelist-based RBAC authorization layer over the knowledge base
with category-level and article-level overrides, audited against UI source of
truth, plus a defense-in-depth link sanitizer that scrubs forbidden article URLs
from media payloads — verified zero leaks across 28 RBAC probe cases (16
allowed, 12 denied).
• Engineered relational people-lookups against the HRIS REST API: direct
phone/email resolution for HR BP and reporting chain, C-suite resolution by
position title (9 roles), and arbitrary employee resolution by name with
Ukrainian morphology tolerance — converting “телефон Ігоря Павленка” into a
structured HRIS query and returning the work phone in under a second.
• Built organizational-intelligence handlers that go beyond static FAQ: newly-hired
employees with configurable look-back window, direct reports enriched with full
profile fetches, and category browsing — turning the bot from a Q&A tool into
an operational assistant for managers.
• Implemented a two-tier caching strategy (TTL employee cache + scheduled
hourly KB refresh via APScheduler) and an audit log of every surfaced article ID
for compliance traceability — enabling post-hoc investigation of any alleged
access violation.
• Designed slug-aware deep links matching the HRIS public-URL scheme via
Cyrillic→Latin transliteration; built UX fallbacks for image-only KB articles (up to
6 inline images with semantic placeholder text) so users always get usable
output even when source content is minimal.

Svidomo — Headless e-commerce backend MAY 2025 — AUGUST 2025
Stack: Python, Django, Django REST Framework, PostgreSQL, Redis, Docker, Nginx
• Architected and developed the backend for a headless e-commerce platform,
enabling integration with any frontend technology, and documented 25+ REST
endpoints (catalog, cart, orders) — accelerating frontend development by ~40%.
• Engineered a stateful cart using Django's session framework, ensuring data
persistence for unauthenticated users.
• Optimized PostgreSQL queries, reducing the average response time for catalog
pages by 200 ms.
• Customized the Django admin panel for product management, allowing content
managers to update 95% of the site's information without developer
intervention.