Maulenov

Rodion Maulenov
AI AUTOMATION & BACKEND ENGINEER
Kyiv, Ukraine · [відкрити контакти](див. вище в блоці «контактна інформація») · [відкрити контакти](див. вище в блоці «контактна інформація»)

SUMMARY

Sole engineer of a ~100k-LOC Python platform that runs a clinic group's operations — six products on one core, LLMs in the
loop, deployed and operated on AWS by me alone. I also run the marketing that platform serves, so I build automation as its
own first user.

CORE SKILLS

Backend Python 3.13, async-first. Django + DRF — where I started and still ship: ORM and admin, server-rendered
templates, SimpleJWT, django-filter, Celery + django-celery-beat, django-storages/S3, gunicorn behind
nginx. Async FastAPI: aiogram 3.x, asyncpg, Pydantic v2, Dishka (DI), uvicorn/uvloop, structlog. SQL,
Bash.
AI / LLM Claude (Agent SDK, CLI process pools, MCP tool servers), Gemini Vision on Vertex AI, OpenCV, prompt &
agent design, structured extraction, LLM-as-judge evaluation, custom multi-agent orchestration written
in-house (in place of LangGraph / CrewAI), cost & rate-limit budgeting
Data & cloud PostgreSQL 17 (schema design, indexing, migrations, event sourcing), Redis, Supabase, pandas,
idempotent ETL, JSONB · AWS (EC2, S3, IAM, registry-based delivery) · GCP (Vertex AI, Cloud Storage)
Frontend Angular — four production apps (Material, CDK, RxJS, reactive forms, Karma specs) on Django REST
backends, shipped on 19 and current with the release train (signals, standalone, zoneless); TypeScript,
vanilla ES2022, Biome/ESLint
Automation & Webhooks, REST APIs, n8n, Telegram Bot API + self-hosted server, MTProto, telephony APIs, Google
infra Sheets API, task queues (Celery, taskiq/Redis) · Docker & Compose (multi-stage, hardened), GitHub
Actions CI/CD, Caddy + TLS, Sentry, PostHog, Prometheus
Quality pytest (~10,000 tests, parallel xdist, per-worker Postgres), mypy strict, ruff, Biome, a purpose-built mock
Telegram API server for end-to-end flows
Marketing Meta & Google Ads run hands-on (~$6.5k/month), lead capture and attribution, funnel instrumentation

EXPERIENCE

Biotexcom — fertility clinic group 2016 – Present

Software Engineer — AI Automation 2021 – Present
Sole engineer and architect of the company's internal automation platform — designed, built, deployed and operated alone.
• Took a chat-and-spreadsheet operation to a ~100k-LOC production platform — six products on one shared core
(FastAPI, aiogram, asyncpg, Redis, Dishka DI) plus a 22k-LOC shared library layer. In production it carries a live call
centre: seven operators, ~6,300 call outcomes a month, 93 % of them recorded automatically from the telephony
webhook with the SIP cause code attached — only the ~450 conversations that need a human decision ever reach a
queue.
• Put an LLM into production, not a demo. ~630 model calls a month across seven daily users: staff type free text and
the model turns it into a structured record, translates for the team, and answers questions about how the business works
— grounded in written documents, not in its own memory. 531 of those are stored as structured records: the model
never writes to the database, it fills a schema, and the schema is machine-checkable. Gemini Vision on Vertex AI reads
forms and scans straight out of Cloud Storage.
• Turned a nine-month client programme into a pipeline instead of a spreadsheet. 700+ clients, each moving through a
scheduled chain — lab analyses, treatment courses, first and second doctor consultations, travel — where every state
change is an append-only event and every deadline a scheduled job. 419 events carry machine-written structure rather
than a manager's prose.
• Run production end to end, alone. GitHub Actions builds a multi-stage image → registry → SSH deploy to AWS EC2
with a health-gated rollout: 1.6 to 2.9 minutes from commit to running container, and the pipeline fails if the container is
not healthy in 60 s. Hardened Compose stack, internal-only network for Postgres and Redis, nightly backups with a
restore-verification script. Integrations I owned end to end: a self-hosted Telegram Bot API server to lift the 20 MB
download cap, an MTProto daemon for what a bot token cannot do, a telephony provider's call API, and MCP tool servers
exposing internal tools to AI agents.
 • Made quality and process mechanical, not manual. ~10,000 tests across 1,222 modules running fully parallel with a
per-worker Postgres clone, a mypy-clean production surface, 59 migrations applied transactionally under an advisory lock,
idempotent pipelines that resume instead of duplicating. Every feature moves through four gated stages — Design, Plan,
Execute, Post-work — and whoever wrote a stage never reviews it: a commit is blocked without a completed self-
verification block, a rule change without the test that guarantees it.

Head of Asian Department — client acquisition & marketing 2016 – Present
Own the Asian market end to end — held in parallel with the engineering role above.
• Built the department from zero: client acquisition, advertising budgets, the communication funnel from first contact to
signed client, and the team that runs it.
• I automate marketing operations I am personally accountable for, and I use the result every day. The manual
workflows I ran here — paid-ad lead capture, outbound calling, the qualification funnel — are the three surfaces I later
rebuilt as automated products.

Enlil, Inc. — MedTech development platform (Campbell, CA) 2022 – 2023

Backend Developer in parallel with the role above
• Built backend services and REST APIs on Django, DRF, Celery and Redis over PostgreSQL for a regulated medical-
device lifecycle platform — FDA 21 CFR Part 820 / Part 11 and ISO 13485 workflows. Containerised with Docker, shipped
through CI/CD on DigitalOcean and AWS.
• Worked inside regulated-software constraints — append-only audit trails, versioned records, traceability from
requirement through verification — in a distributed US/Ukraine team, in English, through code review and written specs.

SELECTED PROJECTS

Autonomous self-repair engine (~20k LOC) — a weekly pipeline that finds stale or unenforced code and rules and repairs
them under limits frozen in Python: four workers, twelve repairs, two attempts, a three-hour wall clock and a spend cap. Each
worker runs in its own git worktree; every fix is judged by an independent zero-context verifier and reverts itself if rejected; three
quality failures in a rolling week trip a breaker that shuts the machine down. The models only propose diffs and verdicts.

Code-reuse mechanism (developer productivity) — a clone radar that fingerprints by content and reds only on NEW
duplication, never on existing debt; a consumer registry generated from the real import graph, so a component's place in the
layering is computed rather than argued; a migration blacklist that fails the build if a retired API reappears; complexity and file-
size ratchets that shrink but never grow.

Real-time call-outcome classification (R&D, validated in production) — a server-side listener joins live calls and Gemini
classifies what it hears, so the system sets the disposition itself. Measured on real traffic: ~87 % capture in the measured
window, 4 simultaneous conversations on one line, ~$0.0015 per classification.

Django + Angular services, Biotexcom (2023 – 2025) — five Django services delivered alongside the main platform: a
regional CRM, a document-extraction service, a scheduling-and-reporting service on Celery Beat with S3, an events API on
DRF with JWT and OpenAPI, and a records backend — each with an Angular application on top. Containerised, most with
GitHub Actions delivery. The repositories belong to the employer; walkthrough available on screen.

EDUCATION & LANGUAGES

Education Bachelor's degree — Dnipropetrovsk Humanities University, 2010 – 2014
Languages Russian — native · Ukrainian — conversational · English — B1 (Intermediate)