M. Gvozdev / Annual report 2026 p. 01 cover

Annual report 2026

p. 01

applied ai engineer

LLM systems, from proof of concept to production

I design, build, and ship production LLM applications for enterprise clients, and I own them end to end: discovery, architecture, proof of concept, deployment, and evaluation. Before AI, I spent twelve years delivering industrial projects for clients.

Now
Applied AI Engineer Consultant, Bitovi
Based in
Montréal, Québec (Eastern Time)
Practice
Applied AI / Software architecture / Software development / Product engineering
Elsewhere
GitHub / LinkedIn / Hugging Face

How to readScroll as usual, or turn the pages: lift the corner of the page, press ← →, or swipe sideways.

Letter

to the reader

p. 03

I came to applied AI from the field. For twelve years, I ran industrial projects in project logistics and specialized rigging: more than 650 of them, with budgets up to $2.3M and teams of up to 30 people across six locations and multiple time zones. That work taught me to scope carefully, to speak plainly with everyone from engineers to executives, and to finish what I start.

In early 2023 I started building with machine learning models full time. Since then I have shipped RAG tools, agentic systems, MCP servers, and evaluation pipelines for enterprise clients, a university research lab, and a startup I co-founded. At Bitovi I own the architecture and delivery of complex AI applications that thousands of people use every day.

I still work the way a project manager does. I start from the problem, prototype the solution, and gather feedback from those who will use it before I start building. Read further for the divisions of that practice, the record of operations, and a few experiments.

Mike Gvozdev Montréal, November 2026

Highlights

highlights of the year

p. 04
  • 6,000

    daily users of a greenfield enterprise AI application, a RAG chatbot with LLM-assisted messaging that I architected and delivered end to end

    Bitovi, 2026
  • 100,000+

    employees at the Fortune 500 company whose internal platform my AI tools are built into

    Bitovi, 2025–26
  • 20–30×

    lower single-call latency, with 2–3× more throughput, after infrastructure upgrades on McGill's IMAGE platform

    McGill SRL, 2025
  • 2.7K

    GitHub stars for claudish-to-english, an open-source Claude Code plugin released in August 2026

    Open source, 2026
  • 126K

    users in one week for a generative AI charity app, with donations up 4×

    Baby Pigeon, 2023–24
  • 95%

    less document processing time after an LLM automation I scoped and shipped for Aplicar

    Baby Pigeon, 2023–24
  • 650+

    industrial projects delivered before software, with budgets up to $2.3M

    2010–2022

Figures as of November 2026

Divisions

four divisions

p. 06

The practice runs as four divisions. Most projects draw on all four at once.

applied ai

Production LLM applications: retrieval-augmented generation, agentic and multi-agent workflows, MCP servers and tools, structured outputs, guardrails, and evaluation pipelines.

CaseRAG chatbot and LLM-assisted messaging for 6,000 daily users
Equipment

Agents: LangGraph, LangChain, OpenAI Agents SDK, Claude Agent SDK, Claude Code, MCP.
Models: fine-tuning with LoRA, QLoRA, and PEFT; quantization with GPTQ, AWQ, and INT8; inference on vLLM; few-shot and chain-of-thought prompting.
Ops: evaluation pipelines, LiteLLM, Langfuse, Weights & Biases, embeddings and vector search.

software architecture

Solution design that is conceived in discovery and implemented in production: how models, data pipelines, APIs, and frontends fit together, and how to make apps performant and scalable.

CaseData architecture and pipelines for an internal platform at the world's largest cruise corporation
Equipment

Design: solutions architecture, ML component integration, data ETL pipelines, user-safety guardrails, SQL and vector databases.
Cloud: GCP (Cloud Run, Vertex AI), AWS (Bedrock, Lambda, EC2, S3, ECS), Azure.

software development

Hands-on build in Python and TypeScript: FastAPI and Flask services, Node.js, REST APIs, SQL, containers, and CI/CD. I write well-structured, maintainable, testable code.

CaseComponents for IMAGE, McGill's open-source accessibility platform, across backend, orchestration, and infrastructure
Equipment

Languages: Python (FastAPI, Flask), TypeScript and JavaScript (Node.js), SQL (PostgreSQL, MySQL).
ML: PyTorch, Transformers, FAISS, Qdrant, Pinecone.
Delivery: Docker, Kubernetes, Terraform, GitHub Actions, CI/CD.

product engineering

From pain point to solution: customer discovery, scoping, proof-of-concept and MVP builds, launches, feedback-driven updates, maintenance, and growth.

CaseBrand new tooling for enterprise executives that replaced a legacy macro-driven workflow
Equipment

Delivery: customer discovery, PoC, MVP, and proof-of-value builds, presales-to-postsales ownership.
Teams: Scrum and Kanban, stakeholder management, cross-functional coordination, mentorship.

Operations

record of operations

p. 08

Applied AI since 2023, built on twelve years of engineering project delivery.

Seven positions, newest first
Nov 2025 – nowApplied AI Engineer ConsultantBitovi, Chicago, IL (remote)AI tools for a Fortune 500 platform, and a greenfield app with 6,000 daily users.
  • Developed and integrated AI-driven tools into the internal platform of a Fortune 500 company with more than 100,000 employees.
  • Owned architecture and delivery of a greenfield enterprise AI application for 6,000 daily users, a RAG-based chatbot with LLM-assisted messaging, end to end in Python from discovery and solution design to production readiness.
  • Designed and deployed agentic workflows, an MCP server and tools, context-engineering utilities, data ETL pipelines, REST APIs, SQL and vector databases, and user-safety guardrails.
Feb – Dec 2025ML Engineer & Solutions ArchitectShared Reality Lab, McGill University, Montréal (hybrid)ML architecture for IMAGE, an open-source multimodal accessibility platform.
  • Defined the ML architecture for IMAGE and its integration across frontend, orchestration, and hardware layers; built multimodal LLM components that interpret visual information.
  • Cut single-call latency 20–30× and improved throughput 2–3× through infrastructure upgrades; raised prediction accuracy through model selection, a revised data pipeline, structured outputs, and prompt engineering.
  • Built a component that converts textbook diagrams into structured formats for visual, audio, and haptic accessibility.
  • Deployed pipelines for model evaluation and selection, prompt testing, and hyperparameter tuning.
Jun – Aug 2025AI EngineerTextlayer, Ottawa (remote)Agentic tools and an evaluation framework for enterprise deployments.
  • Shipped agentic tools for enterprise deployments, including Notion natural-language search and text-to-SQL, and an evaluation framework for response, tool-call, and structured-output accuracy.
Feb – Jul 2025Co-founder & CTOIngsikt, Oslo (remote)Real-time meeting feedback, built on live LLM inference.
  • Designed and prototyped a real-time meeting-feedback product built on live LLM inference, with a recommendation system that tracks key metrics for meeting leaders. Owned product vision, architecture, and delivery.
Jan 2023 – Jan 2025Applied AI EngineerBaby Pigeon Studio, Vancouver (hybrid)Generative AI products and client automation, from scoping to launch.
  • Scoped and shipped an LLM document-automation solution in production for the client Aplicar, cutting processing time by 95%.
  • Built and deployed a RAG chatbot over regulatory and internal documentation for a construction-industry client.
  • Led design, development, and launch of the generative AI products fixgrammar.ai, myEras, and NaiVIGATOR; scaled a charity app to 126K users in one week, growing donations 4×.
  • Created research applications exploring LLM, VLM, and CNN capabilities: Project Cloudberry, Slouch Detector, and AI Rainbolt.
Jan 2020 – Sep 2022Project ManagerApex Specialized Rigging & Moving, Vancouver (hybrid)50+ industrial projects end to end, as the primary client contact.
  • Delivered 50+ industrial projects end to end as the primary client contact, from presales scoping through planning, execution, and closure.
  • Introduced Scrum, cutting planning time by 5–15%.
May 2010 – Sep 2019Lead Project ManagerAcross Project Logistics & Keen-Mark Group, Saint Petersburg600+ projects, budgets up to $2.3M, teams of 10 to 30.
  • Directed full-cycle delivery of 600+ projects with budgets up to $2.3M.
  • Managed engineering teams of 10 to 30 people, ten project teams, and 20+ stakeholders across six locations.

R & D

research & development

p. 11

Side projects and research builds. I like exploring what models can do and how people want to use them.

Featured open source, Aug 2026

claudish-to-english

2.7KGitHub stars
132forks

A Claude Code plugin that shows a plain-English rewrite of each assistant message. The rewrite comes from a local model through ollama by default, or from the codex CLI, the Anthropic API, or any OpenAI-compatible API. It only changes what you read: Claude's reasoning and the saved transcript keep the original, and every hook fails open, so a provider outage never hides an answer.

ClaudishRefactored the hydration boundary so the selector graph memoizes upstream; retries are idempotent now and short-circuit on stale cache keys.
EnglishI changed how the page loads its data so it repeats less work. When a request fails and is retried, it no longer runs twice.
github.com/gvzdv/claudish-to-english ↗
See more research projectsLaboratory notebook5 projects, 2024–2025

  • wiktok

    Jan 2025

    A TikTok-style feed for Wikipedia. DeepSeek-V3 rewrites random articles in internet voices, from greentext biographies to tech-review inventions, and Google TTS narrates them over looping video.

    FastAPI / React / DeepSeek-V3 / Google TTSgithub.com/gvzdv/wiktok ↗

  • naivigator

    Oct 2024

    An AI tour guide. Describe what you want to see, in any words, and it turns the request into places on a map.

    LLM / Maps / JavaScript / Pythongithub.com/gvzdv/naivigator ↗

  • fixgrammar

    Oct 2024

    A writing tool for the job people kept handing to ChatGPT: it corrects spelling, grammar, and punctuation, and nothing else.

    LLM / JavaScriptgithub.com/gvzdv/fixgrammar ↗

  • slouch detector

    Jan 2024

    Webcam pose tracking that watches your ear, shoulder, and knee, keeps score of your posture, and promises boba tea for a straight back.

    Pose estimation / CNN / Pythongithub.com/gvzdv/pose ↗

  • ai rainbolt

    Jan 2024

    A vision model plays GeoGuessr. Press space and it screenshots the round, makes its guess, and says it out loud through ElevenLabs.

    Vision LLM / OpenAI / ElevenLabsgithub.com/gvzdv/ai-rainbolt ↗

Personnel

personnel file

p. 13
Education & certification
Claude Certified Architect
Anthropic, 2026
MSc, Engineering Management, with honours
Saint Petersburg State University of Architecture and Civil Engineering, 2010
Professional Scrum Master I
Scrum.org, 2022
Bullet Proof Manager leadership program
Crestcom International, 2022
Languages
  • englishFluent
  • русскийNative
  • norskIntermediate
  • françaisLearning

Nitsche's General Dynamics posters were printed in six languages. Counting generously, I'm at three.

Correspondence

correspondence

p. 14

Write about a project, a role, or a problem you'd like a second pair of eyes on.

ping@mikegvoz.dev

Set in Saira, Libre Franklin, and IBM Plex Mono. The plates are drawn in code after Erik Nitsche's posters and reports for General Dynamics, 1955–1960.

Printed in Montréal, 2026