Богдан

BOHDAN HAVRYLIUK
AI Engineer · LLM Systems · Agents · Python Backend
[відкрити контакти](див. вище в блоці «контактна інформація») | [відкрити контакти](див. вище в блоці «контактна інформація») | Leicester, UK | [відкрити контакти](див. вище в блоці «контактна інформація») |
github.com/bogdan734 | kalorad.app
Available part-time now (up to 20 h/week in term, more in vacations) and full-time from 2027. Remote across
England or hybrid in Edinburgh/Scotland. Right to work: UK Student visa, Graduate route from 2027.
SUMMARY
AI engineer with 5+ years in software and 4+ in applied AI. Led from-scratch LLM pretraining up to 8B
parameters (150B tokens) on a EuroHPC grant; ship production LLM systems: a voice recruiter with 300+ live
calls and the AI layer of a P2P payments platform (sole AI developer, middle backend). Python, FastAPI,
ASP.NET Core; evaluation-driven. Final-year BSc Artificial Intelligence, De Montfort University (first-class
average 71.7%).
CORE SKILLS
LLM & ML: LLM pretraining, SFT/LoRA, GRPO, RAG, agents and tool use, embeddings, evaluation and
benchmarking, PyTorch, FSDP, SLURM/HPC, computer vision (YOLO, OCR), transfer learning, Claude API,
prompt engineering
Backend: Python (FastAPI, asyncio, SQLAlchemy, Pydantic, pytest), C# / ASP.NET Core, TypeScript (React,
Next.js, NestJS), SQL, REST, webhooks, PostgreSQL, SQLite, Redis, Docker, Linux, Git, CI
Integrations: Telegram (Telethon, bots), Vapi, Deepgram, ElevenLabs, SIP, Stripe, CRM APIs (KeyCRM)
Languages: English (fluent), Ukrainian and Russian (native). Also: C++, Swift, PHP, Lua
EXPERIENCE
Founder & Lead Engineer — Kalorad May 2026 – Present
Own product; an engineer has joined, a second hire is in progress
Led from-scratch pretraining of three language models: 683M (13.1B tokens, about $86), a 1.3B-active
mixture-of-experts (60B tokens) and an 8B dense model (150B tokens, finished Oct 2026), the last two on a
EuroHPC grant of 10,000 H100 GPU-hours at MareNostrum 5; distributed training with FSDP on SLURM.
Built the data and evaluation stack: quality filtering, MinHash deduplication, benchmark decontamination and
a harness reporting 95% intervals; identified and fixed 9 training bugs, each covered by a regression test.
Designed and shipped the product on ASP.NET Core 9: streaming chat, RAG with cite-or-abstain, tool router,
20 agent packs, MCP connectors, Telegram bots, an OpenAI-compatible API, and agent bots with memory,
scheduled routines, teams and learned skills running in production.
Trained an 80M multilingual embedding model (recall@1 0.80) and built a GRPO pipeline on sandbox-
verified tasks; covered the system with 485+ automated checks.
Audited the 97k-line codebase (177 findings), fixed CSRF, a stored XSS and vulnerable dependencies, and
added CSP, rate limiting, TOTP two-factor login and Stripe billing.
Middle Backend & AI Developer — Payment-processing company (P2P payments) Sep 2026 – Present
Sole AI developer
Built an AI layer over Telegram support groups (40,850 messages, 126 groups): Telethon ingestion, PII
masking before any external API, and a rules-first pipeline replayed in shadow mode against human
answers.
Measured that 49% of client turns waited over a 15-minute SLA; built an intent taxonomy (922 labelled
items; at first release 89% of merchant requests matched, 91% and 87% precision on a 510-message
sample), an SLA watchdog and a live listener that reconciles after downtime.
Added a labelling bot with roles and agreement reports and local OCR of payment receipts for Ukrainian and
Turkish banks; compared a learned decision head and a local LLM with the rules. The system never moves
money.
AI Engineer / AI Programmer (contract) — STL Aug 2026 – Present
Part-time
Building a computer-vision system that finds and reads licence plates in vehicle photos, matches a queried
plate against the photo set and returns ranked candidates (top-10), with make/model classification.
AI Trainer — Meta; AI Training Contractor — Scale AI Jan 2025 – Jul 2026
Part-time contracts
Rated and compared model outputs on coding, algorithm and reasoning tasks for correctness, safety,
complexity and security; solved advanced Python exercises to expose failure modes and flagged policy-
violating responses.

Team Lead & Software Developer — client under NDA Sep 2023 – Sep 2025
Led project teams, architected backend and full-stack applications in C#, .NET, Python and SQL, and
enforced code-quality standards.
Freelance Full Stack AI Engineer — private clients 2021 – Present
Delivered full-stack, backend and AI automation projects end to end (Python, C#/.NET, Node.js,
PostgreSQL, Docker).
SELECTED PROJECTS
EVA — voice AI recruiter (Python, FastAPI, Claude API, Vapi, PostgreSQL · github.com/bogdan734/eva-ai-
recruiter). Sole developer of a production system that calls candidates in Ukrainian, runs structured screening
and routes qualified people to recruiters (300+ live calls); end-to-end KeyCRM integration, a Telegram follow-
up agent, SIP diagnostics, voicemail detection and self-healing reconciliation.
CCTV Video Analytics (Python, YOLO11, ByteTrack · github.com/bogdan734/cv-video-analytics). Two
pipelines on a shared detection and tracking core: finds the moment two vehicles collide (one event, no false
positives in the remaining 64 s) and logs each child appearance exactly once (66 appearances, up to 27
people in frame).
AI Operations & Follow-up Service (Python, Anthropic API, SQLite · github.com/bogdan734/techcom-ai-
ops). Rules decide who needs a nudge, an LLM drafts the message, an idempotent outbox sends each follow-
up once (a second run sends 0 messages and makes 0 model calls); handles API and model outages and
prompt injection; 218 tests.
Nexus — AI study workspace (Next.js, NestJS, Prisma, Electron · github.com/bogdan734/Nexus). 11 LMS
integrations through official APIs only, TOTP two-factor login, AES-256-GCM encrypted credentials, an SSRF
guard and a desktop client with device binding.
Konspekt and Telegram Mini Games (Swift 6, TypeScript, React · github.com/bogdan734/konspekt-miniapp ·
github.com/bogdan734/tg-mini-games). A native iOS app that turns recorded lectures into study notes, a
Telegram Mini App with spaced-repetition flashcards and tests, and a Mini App hub of two tower-defence
games with unit-tested engines.
Kozyr Trans landing page (in progress). Rebuilding a freight carrier's landing page to replace a 703 KB Tilda
page: route and weight calculator, fleet picker, a Python lead backend with pytest tests and automated site
audits.
AI freight-forwarding CRM (personal project, planning stage). Architecture and 3-month plan for AI agents
that take a request to agreed terms, with swappable channel, LLM and telephony adapters.
EDUCATION
BSc (Hons) Artificial Intelligence — De Montfort University, Leicester 2023 – 2027
Current average: First Class (71.7%). Modules: Machine Learning, Neural Networks, Computer Vision,
Intelligent Systems, Data Structures & Algorithms, Software Engineering. Project: transfer learning (AlexNet,
GoogLeNet) for five-class rice classification with augmentation and 5-fold cross-validation.