Богдан

BOHDAN HAVRYLIUK
Full Stack AI Engineer • AI Automation • Backend • LLM Systems
[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
Profile
Full Stack AI Engineer with 5+ years of software engineering experience and 4+ years of hands-on AI development, model evaluation, and AI-enabled product delivery. Experienced in building end-to-end systems across async Python, .NET, LLM integrations, backend APIs, databases, cloud deployment, and voice AI. Built and support production automation used by real recruiters and candidates. Currently founder and sole engineer of Kalorad, a self-hosted AI platform running on language models I trained from scratch (up to 8B parameters). Final-year BSc Artificial Intelligence student at De Montfort University, Leicester (Student visa, up to 20 hours/week during term); available full-time from 2027. Seeking remote roles across England or hybrid in Edinburgh/Scotland.
Key Skills
• Languages: Python, C#, C++, C, TypeScript, JavaScript, Swift, PHP, Lua, SQL, HTML/CSS
• AI / LLM: Claude API, LLM tool use, prompt engineering, AI response evaluation, RLHF / AI training pipelines, NLP, computer vision, voice AI, PyTorch, LLM pretraining & fine-tuning, self-hosted inference, RAG, GRPO, distributed training (FSDP, SLURM, HPC), embedding models
• Backend & Infrastructure: FastAPI, asyncio, SQLAlchemy 2, Alembic, Pydantic, .NET / ASP.NET, Node.js, PostgreSQL, SQLite, Redis, Docker, Docker Compose, Caddy, VPS
• Integrations: Vapi, Deepgram, ElevenLabs, SIP, Telegram MTProto / Telethon, KeyCRM REST, webhooks, REST APIs, S3-compatible storage
• Languages Spoken: English (fluent), Ukrainian (native), Russian (native)
Experience
Founder & AI Engineer — Kalorad
May 2026 – Present (Own product)
• Trained language models from scratch as the sole engineer: a 683M model (13.1B tokens, about 43 GPU-hours, about $86), a 1.3B-active mixture-of-experts model (60B tokens, 57,221 steps) and an 8B dense model (150B tokens, 71,526 steps, finished 25 Oct 2026), the last two on a EuroHPC grant of 10,000 H100 GPU-hours at MareNostrum 5 (approved Oct 2026). Built the data pipeline (quality filtering, MinHash deduplication, benchmark decontamination), distributed training on SLURM with FSDP, and an evaluation harness reporting 95% intervals; found and fixed 9 training bugs, each with a regression test.
• Developed the platform on ASP.NET Core 9: streaming chat, RAG with cite-or-abstain, decaying long-term memory, a permissioned tool router, 20 vertical agent packs, connectors and MCP support, Telegram bots, desktop, browser and Office client apps, and an OpenAI-compatible public API.
• Built agent bots with persistent memory, scheduled and webhook-triggered routines, multi-agent teams and learned skills, running in production with sandbox isolation and an outbound-message sentinel; Slack, Discord, WhatsApp, email and A2A are implemented behind feature flags.
• Trained an 80M multilingual embedding model (recall@1 0.80 on held-out pairs) and implemented an LM-based LoRA embedder with Matryoshka and hard-negative mining, plus a GRPO pipeline that trains on model-written, sandbox-verified code and maths tasks.
• Implemented Stripe billing (subscriptions, one-off SKUs, usage metering), a referral loop with anti-abuse trust levels, push chains and an admin analytics dashboard.
• Audited my own 97k-line codebase (177 findings, with a fix plan) and fixed real vulnerabilities: CSRF across 33 of 37 POST endpoints, a stored XSS, a plugin-CSS script-injection path and critical vulnerable NuGet dependencies. Added a site-wide CSP, rate limiting, account lockout, TOTP two-factor login, email verification, password reset and session revocation.
• Covered the system with 485+ executable smoke and end-to-end checks (web, API, agents, bots, privacy, config) run on every build.
• Secured a EuroHPC AI Factories access grant (10,000 H100 GPU-hours at MareNostrum 5, approved Oct 2026); prepared investor and grant materials (pitch decks, technical overview, financial model) and applications to programmes including EBRD Star Venture and Hub71.
AI Engineer / AI Programmer (contract) — STL
Aug 2026 – Present (Part-time freelance, contract basis)
• Engaged as a contracted specialist (official contract title: Engineer-Programmer / AI Engineer / AI Programmer) to develop a computer-vision system for vehicle identification from photographs.
• Building a detection and OCR pipeline that scans vehicle photos, locates and reads license plates, and matches a queried plate number against the indexed photo set, returning ranked candidate matches (e.g. top-10).
• Added automatic make/model classification for matched vehicles.
• Delivered as an independent, part-time contract engagement alongside other work, rather than in-house full-time employment.
AI Training Contractor — Scale AI
Jan 2026 – Jul 2026 (Part-time freelance, Remote)
• Evaluated AI-generated code and technical responses for correctness, helpfulness, safety, and best practices.
• Performed side-by-side comparisons across coding, algorithm analysis, and problem-solving tasks.
• Flagged harmful or policy-violating responses to support model quality and safety improvements.
AI Trainer — Meta
Jan 2025 – Feb 2026 (Part-time contract, Hybrid)
• Contributed to training and fine-tuning large-scale AI models through structured data generation and feedback.
• Reviewed and rated model outputs across technical and general domains to improve response quality.
• Applied software development and AI expertise to assess accuracy, usefulness, and clarity of model content.
Team Lead & Software Developer — Client under NDA
Sep 2023 – Sep 2025 (2 Years)
• Led project teams to deliver complex software and AI solutions on time and within scope.
• Architected and built backend and full-stack applications using C#, .NET, Python, and SQL.
• Enforced code-quality standards and drove continuous performance optimisation.
Freelance Full Stack AI Engineer — Self-employed
2021 – Present (ongoing private-client work, alongside the roles above)
• Designed and delivered full-stack, backend, and AI automation solutions for private clients, from requirements and UX through development, delivery, and support.
• Built production applications using Python, C#, .NET / ASP.NET, JavaScript / Node.js, SQL and PostgreSQL, REST APIs, Docker, and responsive UI/UX workflows.
• Built scalable business logic, integrated relational databases, and maintained high code standards across diverse technology stacks.
• Delivered several non-disclosable professional projects under NDA.
Production Projects
AI Voice Recruiter "EVA" — End-to-End Hiring Automation (Python, FastAPI, Claude API, Vapi, PostgreSQL; github.com/bogdan734/eva-ai-recruiter)
• Sole developer of a production system for Kozyr Trans that autonomously calls candidates in Ukrainian, runs structured screening, and routes qualified candidates to recruiters; processed 300+ live calls.
• Integrated KeyCRM end to end: candidate cards are created and moved through the funnel automatically with transcript, call recording, AI summary, and structured screening fields.
• Built a Telegram MTProto follow-up agent for unreachable candidates, merging voice and chat dialogs into the same CRM card with deduplication.
• Implemented async services with FastAPI, SQLAlchemy 2, Alembic, PostgreSQL, Claude API, Vapi, Deepgram, ElevenLabs, SIP, webhooks, Docker Compose, Caddy, and APScheduler.
• AI Support Automation for a Payment-Processing Company
AI Support Automation for a Payment-Processing Company (Python, Telethon, pytest) (Sep 2026 – Present)
• Built a Telegram ingestion and analytics pipeline over 40,850 messages in 87 merchant and trader support groups (60 days); measured that 49% of client turns waited longer than a 15-minute SLA (median first reply 13 min, p90 3 h).
• Designed an intent taxonomy and rule engine (89% of merchant requests classified automatically; 91% / 87% precision on 510 hand-labelled messages), an SLA watchdog, shadow-mode replay comparing rule replies with human answers, and PII masking for all data leaving the machine.
AI Freight-Forwarding CRM (personal project; architecture and delivery plan, planning stage)
• Designed a 3-month plan for a custom CRM in which AI agents take a request from first message to agreed terms with both sides, contact a carrier base of thousands over Viber, Telegram and WhatsApp, and leave managers only payment and document checks.
• Specified hot-swappable ports/adapters for channels, LLM and telephony, a Telegram Mini App, and analytics (conversion, ROI, average check, AI spend).
Kozyr Trans — Corporate Website / Landing Page Redesign (in progress)
• Rebuilding the public landing page of Kozyr Trans, a Ukrainian freight carrier, to replace a 703 KB Tilda page: hero route/weight calculator, fleet picker, case photos and a scroll-driven route map.
• Python backend routes every lead through a single delivery seam; AI screening questions on the thank-you page feed the lead card; pytest tests and automated site audits.
Selected Full-Stack Projects
Nexus — AI Study Workspace for Students (Next.js, NestJS, Prisma, PostgreSQL, Electron; github.com/bogdan734/Nexus)
• Built a full-stack study platform with 11 LMS integrations (Canvas, Moodle, Brightspace and others) through official APIs only, aggregated deadlines, and student-owned drafts with version history; auto-submission for credit is disabled by design.
• Implemented TOTP two-factor login with recovery codes, AES-256-GCM encrypted credentials, an SSRF guard for user-supplied URLs, device binding by hardware id, and an Electron desktop client with an OS-keychain token vault.
dZENcode Comments SPA (Django, DRF, Channels, Celery, RabbitMQ, PostgreSQL, Redis, Elasticsearch, Vue 3)
• Built a Dockerized cascading-comments SPA with unlimited reply depth, CAPTCHA, attachments, JWT auth, real-time WebSocket updates, live preview, validation, and XSS/SQLi protections.
• Designed scale-ready backend features: GraphQL, async search indexing, idempotency middleware, rate limiting, S3/CDN attachment support, Prometheus metrics, read-replica routing, partitioning template, and Locust load tests.
AI Operations & Follow-up Service (Python, Anthropic API, SQLite; github.com/bogdan734/techcom-ai-ops)
• Built a service where deterministic rules decide which project tasks need a nudge, an LLM reads ambiguous notes and drafts the message, and an idempotent SQLite outbox sends each follow-up exactly once: a second run on the same data sends 0 messages and makes 0 model calls.
• Handled flaky APIs, provider and model outages and prompt injection with bounded retries, a circuit breaker and template fallbacks; 218 tests run in under a second with no network.
CCTV Video Analytics — Collision and Child Detection (Python, YOLO11, ByteTrack/BoT-SORT; github.com/bogdan734/cv-video-analytics)
• Built two pipelines on a shared detection and tracking core: one finds the moment two vehicles collide (one event at 00:30.000, no false positives in the remaining 64 s), the other counts people and logs each child appearance exactly once (66 appearances, up to 27 people in frame).
• Added tests, CI and run artefacts (annotated video, log, JSONL, CSV) with a requirements-coverage table.
Konspekt — Study Apps (Swift 6, iOS; Telegram Mini App; github.com/bogdan734/konspekt-miniapp)
• Built a native iOS app that records lectures and turns them into structured study notes, with a tested Swift package and a content pipeline covering anatomy, chemistry, Latin, English and history.
• Also shipped a static Telegram Mini App with syllabus, timetable, notes, spaced-repetition flashcards, tests and homework, with progress kept on the device and no server or keys.
University Projects
Rice Variety Classification via Transfer Learning (MATLAB, AlexNet / GoogLeNet)
• Applied transfer learning, data augmentation, stratified splits, SGDM fine-tuning, and accuracy/confusion-matrix evaluation.
Education
BSc (Hons) Artificial Intelligence — De Montfort University, Leicester
2023 – 2027 (final year)
• Core modules: Machine Learning, Neural Networks, Computer Vision, Intelligent Systems, Data Structures & Algorithms, Software Engineering.
• Current Average: First Class Honours (71.7%).