Денис

DENIS SIRICHENKO
AI Automation Engineer · Python / LLM Backend
Cherkasy, Ukraine — remote, open to EU relocation · [відкрити контакти](див. вище в блоці «контактна інформація») · Telegram: @sordens_work
sordens.com · github.com/YOUR-HANDLE · Ukrainian & Russian native · English B1 (working proficiency)

SUMMARY
Automation engineer who ships production LLM systems end to end — data pipelines, retrieval, agent orchestration, and the
infrastructure underneath. Sole technical owner of a Dubai brokerage’s platform: 159,000+ live listings refreshed daily, 1.7M
government transaction records, three production voice agents and a WhatsApp assistant used by working agents. Building
production AI systems since 2025. Comfortable where the hard parts are: API ceilings, traffic and token budgets, silent failures, and
regulation that must be enforced in code.
TECHNICAL SKILLS
AI & LLM: OpenAI, Anthropic Claude, Google Gemini, OpenRouter, Retell AI voice agents, RAG, pgvector, Pinecone, bge-m3
embeddings, hybrid retrieval (RRF), tool-calling agents, prompt engineering, LLM A/B evaluation harnesses, self-hosted Ollama
(qwen2.5), Whisper / AssemblyAI STT.
Backend: Python — FastAPI, async SQLAlchemy, aiogram, Pandas, Playwright. REST APIs, webhooks, WebSockets, background
workers, session auth, CSRF protection, rate limiting, RBAC.
Data: PostgreSQL at 2M+ rows — schema design, indexing and query optimisation, upserts, integrity constraints, deduplication,
price-history triggers. Alembic migrations, live production DB migration, Redis, SQLite, ETL pipelines, data-quality auditing.
Data acquisition: Resilient scraping with Playwright and direct server-JSON parsing, proxy-pool management (Webshare API,
rotation, IP auth, freshness windows), rate limiting, portal-limit compliance.
Automation: n8n (expert, self-hosted — sub-workflows, tool-calling agents, infrastructure-as-code deployment), Make, ManyChat,
Zapier, Meta WhatsApp Cloud API, Telegram Bot API, Google APIs.
Infrastructure & DevOps: Docker & Compose, Hetzner VPS, Traefik, Caddy, Let’s Encrypt, systemd services, zero-downtime deploys,
Cloudflare, Linux hardening (UFW, Fail2ban, ED25519 SSH), Git / GitHub.
Security & compliance: Password hashing, PII filtering, audit logging, secrets handling, regulation encoded as code (UAE Cabinet
Resolution 56/2024, DNCR screening).
AI-assisted development: Claude Code, Cursor, Codex — used as executors under my architecture and review, not as a substitute
for it.
PROFESSIONAL EXPERIENCE
Freelance AI Automation Engineer | Sordens — independent engineering practice | Sep 2025 – Present
Production LLM systems, data pipelines and self-hosted infrastructure for agency and SMB clients — architecture, build, deployment
and operation.
Key engagement — Smart Start Properties, licensed real-estate brokerage, Dubai, UAE (remote) | 2026
Sole technical owner of the data platform, internal products, AI agents and infrastructure, working directly with the owner, agents
and telesales.
Data platform & ETL
• Built and operated the agency’s property data platform in PostgreSQL: 159,000+ active listings refreshed daily from 3 live
portals, plus 1.7M Dubai Land Department transactions, the Ejari rental registry and the service-charge index.
• Beat a hard 1,250-result API ceiling with recursive query-space partitioning across three axes (price / bedrooms / area), auto-
generating 1,355 ranges from a lightweight volume probe — full catalogue coverage where a single query returned a fraction of
it.
• Cut scraper traffic 36× by dropping the browser: found the full listing set in the page’s server-rendered Next.js JSON, replaced
Playwright with a plain HTTP GET — 10 MB → 278 KB per page, seconds → 0.7 s, 78 extracted fields instead of the previous
handful. Ended a 268 GB overrun on a 250 GB plan.
• Sharded a nightly crawl that ran 21 hours without finishing and overlapped the next run — 1/7 of the catalogue per night, full
coverage weekly, no collisions. Orchestrated over SSH with a per-run Telegram report.
• Ran a data-quality audit that traced Abu Dhabi / Sharjah / Ajman listings leaking into the Dubai database to a search URL built
without a location filter; patched all 5 parsers and removed 2,639 bad records with a verified backup taken first.
Voice AI & telephony
• Shipped 3 production voice agents on Retell AI (sale / rent / cold outreach, EN–RU–AR context) with a custom Python
dispatcher instead of no-code: export filtering, deduplication, DNC lists, call pacing, dry-run by default, live calls only on double
confirmation, Telegram + CSV reporting.
• Encoded regulation directly in the dialer: UAE Cabinet Resolution 56/2024 — a hard 09:00–18:00 Dubai calling window, DNCR
screening and a full audit log. Violations carry penalties up to AED 150,000.
• Diagnosed ~80 calls that connected but returned silence: proved by A/B routing the identical stack through a +380 number
(clean two-way audio) that UAE carriers were filtering third-party VoIP — a network-level cause, not configuration. Saved
weeks of codec tuning. Evaluated 6 telephony vendors by direct support correspondence, scored on price, lead time and test
results.

• Tuned prompts against live calls: temperature 0, prompt budget under 4,000 tokens, English prompts instead of Cyrillic (2–3×
cheaper per token), interruption sensitivity 0.5 with a 700 ms delay after 0.85 shredded speech, semantic prohibitions instead
of word blacklists.
Internal products
• FindObject — internal telesales service on FastAPI + aiogram + async SQLAlchemy + PostgreSQL, deployed with Docker, Traefik
and systemd: web interface and Telegram bot, password auth, Excel / CSV / PDF import, owner lookup. Services split under
systemd, with an internal refresh service bound to loopback behind a shared secret.
• Chose deterministic parameterised SQL with the LLM only classifying query intent over text2sql — predictable results,
predictable cost, no injection surface.
• Unified lead status keyed on the normalised phone number rather than record id, so one owner arriving in several exports
resolves to one lead; added a full “who marked what, when” history table.
• WhatsApp assistant for agents in self-hosted n8n — 26+ nodes, sub-workflows and a tool-calling agent: database search,
branded PDF brochure generation, ROI calculator, DLD-based price estimates, voice input.
• Infrastructure-as-code for n8n: workflow deploy scripts that patch SQLite directly on a stopped instance to preserve the
WhatsApp trigger’s webhookId (regenerated on every UI save, silently killing the Meta callback) and credential bindings.
Required reverse-engineering that n8n executes the published snapshot in workflow_history, not the working copy, and that
active workflows live in memory until container restart.
Reliability, cost & evaluation
• Built a 96-run A/B evaluation harness (16 scenarios × 3 runs × 2 models) with a manual scoring rubric to choose the production
model — winner averaged 9.3 vs 7.8, and was faster and more concise.
• Cost engineering: API response cache with a 24-hour TTL plus a request counter with alert thresholds, keeping the whole
agency inside a 200-request/month free tier; token budgets on every prompt; proxy traffic under plan.
• Production safeguards: guarded sweeps that refuse to deactivate stock when under 80% of the catalogue was collected, a
backup before every destructive DELETE, per-run Telegram reports and crash reports.
Other client work
• Built a NocoDB + n8n + Telegram operations automation for a Ukrainian cleaning company operating in Germany: order intake,
scheduling and staff notifications.
• Deliver end-to-end AI funnels for SMB clients — lead capture → qualification → CRM — with the full stack owned per delivery:
Hetzner VPS, Docker, Caddy / Traefik, Cloudflare, TLS, monitoring and hardening.
AI Automation Mentor & Lecturer | GO IT Education (Ukraine, remote) | Jul 2026 – Present
• Teach live 90-minute practical sessions on LLM APIs, n8n, Telegram AI bots, RAG and multi-tool agents; audience averages ~30
years old, a large share with no technical background.
• Author complete module packages — slide decks, mentor scripts and working n8n workflow JSONs covering OpenRouter,
AssemblyAI transcription with polling, sub-workflows used as agent tools and Google Drive as a knowledge base.
CNC & Laser Machine Operator | Manufacturing | 2024 – 2025
Operated precision automated production systems under strict tolerance and QA control — where the reliability-first habits above
come from.
SELECTED PROJECTS
AI Perfume Consultant — production Telegram bot over ~2,500 perfumes. PostgreSQL + pgvector with a two-vector architecture
(technical attributes and “vibe”), bge-m3 embeddings, RRF hybrid retrieval, self-hosted Ollama (qwen2.5:7b) for review
summarisation and Gemini for query reformulation. Iterated to ~9.5/10 on a manual relevance rubric.
Stack: Python, FastAPI, PostgreSQL/pgvector, Ollama, Gemini, Telegram Bot API
@de_svitlo_che_bot — Power Outage Monitor — Telegram bot tracking live electricity schedules for Cherkasy residents. 500+
organic active users, running unattended in production. Open source on GitHub.
Stack: Python, aiogram, Docker, Hetzner VPS, SQLite
E-commerce catalogue cleanup & semantic search — inherited a raw 10K+ record database; deduplicated 40%+, standardised
fields, built an embedding pipeline and a Telegram bot for fuzzy and semantic product lookup.
Stack: Python, PostgreSQL, vector embeddings, Telegram API
Also: multimodal calorie tracker (voice/photo → macros, Whisper + Vision), autonomous job-scout pipeline, Gmail digest bot with
LLM filtering, KidControl parental-control architecture (React Native + FastAPI + Redis/Celery), self-hosted FastAPI video service.
EDUCATION
B.Sc. Automation & Computer Systems — University, Cherkasy | 2025 – present (part-time, distance)
Junior Specialist, CNC & Robotics — Technical College, Cherkasy | 2021 – 2025