Stepan Fedyanov
Applied AI Engineer

May 2026 – Present

A company whose product portfolio includes Madison AI, an AI platform for US local governments: procurement document generation, staff reports and compliance checks for cities and counties (Merced, Contra Costa County, Corona), plus integrations with procurement marketplaces (BidNet, OregonBuys, ePro).

• Built Solicitation Builder from scratch in collaboration with Sovra, an AI tool that generates procurement documents and exports them to the BidNet platform: architecture, a Claude-based AI pipeline with chat history in Cosmos DB, MySQL with automated migrations in CI/CD, Microsoft Entra ID authorization, and dev, staging and production environments on Azure.
• Developed a Word-like WYSIWYG editor on Tiptap with pagination and print-exact rendering: headers and footers, tables, Edit with AI, Format Standards (agency styles and logo), and PDF/DOCX export through a config-driven exporter. Later moved the editor and the template selection engine into the core Madison AI app.
• Implemented the AI Evaluator, which checks documents against procurement policies: a single Compliance Report and review rail built on one evaluation, citations linked to source documents, anchored Fix with AI edits, and a template language checker.
• Improved the procurement agents: guided chat with one question per turn, correct tool-calling order, chat title and filename generation (with a regex fallback), and a Haiku agent for dates and METS fields. Split chat and document generation in Solicitation Builder into separate models and processes, and cut the Sovra search agent's response time by about a third on average.
• Built data acquisition pipelines for OregonBuys and ePro (Sovra): Azure VMs, Automation Account runbooks, fault-tolerant sync with retry logic, and migration of 13,000+ records from Redshift and S3 to Azure.
• Implemented the Document API (Export-to-METS) for JWT-secured document exchange with BidNet, and a Claude-powered HTML report generator for GFOA with PDF rendering.
• Shipped the public Policy & Law Assistant widget (inline and floating embeds) with per-jurisdiction configuration and a public-path security check script, plus an agent skill that generates new procurement workflows from a template.
• Cut Azure infrastructure costs roughly in half (Defender for Cloud, App Service tiers, redundant database instances), set up monitoring for every environment and Dependabot, fixed vulnerabilities across apps and VMs, and sped up backend tests from 40 to 5 minutes.

Stack: Python, TypeScript, React, Tiptap, Claude API, Azure (App Service, Container Apps, MySQL, Cosmos DB, Blob Storage, Entra ID), AWS (S3, Redshift), Docker, GitHub Actions.

March 2024 – May 2026

The largest gold mining enterprise in Siberia, operating multiple distributed production sites with complex logistics and data flows (2,000+ employees).

• Designed and implemented a full-stack ERP system for gold inventory tracking and operational analytics (250+ users), replacing manual Excel workflows and reducing reporting time from 2 hours to 30 minutes.
• Optimized large-scale PostgreSQL data processing by rewriting critical queries, adding composite indexes, and introducing materialized views, reducing executive report generation time from 50 minutes to under 2 minutes.
• Built a distributed data synchronization layer (Celery, REST) for geographically isolated production sites with unstable connectivity, ensuring reliable and idempotent data delivery across all nodes.
• Designed a multi-agent LLM analytics system (LangGraph, 3 agents) that converts natural language requests into fully rendered HTML dashboards. The system includes a query understanding agent (maps intent to a structured dashboard spec), a data retrieval agent (generates and validates SQL), and a dashboard builder agent (renders HTML/JS output), enabling non-technical managers to access insights without analyst support.
• Implemented a query understanding pipeline that enriches prompts with RAG-retrieved domain context (business rules, metric definitions) and uses schema-constrained outputs (Pydantic) to extract entities, time ranges, chart types, and KPIs into a validated JSON format, ensuring deterministic downstream execution.
• Built a retrieval-augmented context layer by chunking internal documentation, business rules, and DB schema, generating embeddings, and storing them in ChromaDB. At query time, the system retrieves top relevant chunks via semantic search to ground LLM responses and reduce hallucinations in metric definitions and schema usage.
• Developed an automated pipeline that parses Django migration files to extract and normalize database schema metadata into structured JSON, keeping LLM context up-to-date and improving accuracy of text-to-SQL generation.
• Implemented a safe text-to-SQL pipeline with multiple validation layers: LLM-based checks for schema consistency, static SQL validation, and EXPLAIN dry-runs before execution. Combined with read-only access, table whitelisting, and retry logic, this reduced invalid queries and hallucination-driven errors to near zero in production.

Stack: Vue 3, TypeScript, Python, Django/DRF, PostgreSQL, Celery, LangGraph, LLM (OpenAI/Claude/Ollama), RAG (ChromaDB).

May 2024 – May 2026

The #1 design and technology company in Eastern Europe, known for complex digital products and pioneering AI solutions (Fokus, Nikolay Ironov).

• Designed core features for Fokus (AI-powered presentation builder with 100k+ users), including interactive canvas elements and Stripe-based monetization flows.
• Designed and built a distributed LLM-powered presentation generation pipeline that converts a topic, raw text, or URL into fully structured slide decks. Implemented preprocessing for web content (scraping + text extraction) and unified all inputs into a single generation flow.
• Implemented a plan generation stage (Celery) that produces a structured presentation outline with slide types and content descriptions. Injected available slide templates directly into the prompt and enforced strict Pydantic validation with automatic retries, ensuring deterministic and valid outputs.
• Built a parallel slide generation system using Celery groups and chords, where each slide is generated independently. Each task fills template fields via LLM and generates image prompts, which are resolved via external APIs. Isolated failures per slide with retry logic, preventing full pipeline breakdown.
• Designed the final aggregation stage (Celery chord callback) that assembles all slides into a validated JSON structure and persists it to the database, including integrity checks before writing.
• Designed and built an AI-powered code review pipeline that analyzes Git diffs and provides context-aware feedback directly in GitLab Merge Requests based on labels and change scope.
• Implemented a preprocessing layer that parses diffs, filters noise, and enriches context with related files and metadata, allowing the LLM to reason not only about isolated changes but about their impact on the surrounding codebase.
• Designed a prompt pipeline that adapts review depth dynamically (style, architecture, logic) based on MR labels and file types, ensuring relevant and targeted feedback instead of generic comments.
• Built a structured feedback system where LLM outputs are normalized into categorized comments (bugs, risks, improvements, style), enabling consistent formatting and easy consumption by developers.
• Integrated the system into GitLab CI/CD as an automated step, posting inline comments and summaries directly into Merge Requests, reducing review turnaround time by 67% and catching logic issues before human review.
• Led migration from Vue 2 (Nuxt 2) to Vue 3 (Nuxt 3), reducing technical debt and improving maintainability.
• Implemented automated testing (Playwright, Vitest) and monitoring (Sentry), increasing crash-free sessions from 89% to 98%.

Stack: Vue 2/3, Nuxt 2/3, TypeScript, Pinia, Node.js, OpenAI/Gemini, Python (Django/DRF), PostgreSQL, Playwright, Vitest, GitLab CI/CD, FSD.

June 2023 — May 2024

• Developed key generative and front-end features for Nikolay Ironov, the pioneering AI identity generator for small businesses, scaling the product to over thousands of unique design projects.
• Led a major performance refactoring initiative, converting the frontend to leverage Server-Side Rendering (SSR) and eliminating outdated dependencies. This effort improved Google PageSpeed scores from 40 to 80 and drove a ~75% growth in organic search traffic.
• Introduced and enforced the Feature-Sliced Design (FSD) architectural pattern into a large legacy codebase, which significantly increased code clarity, modularity, and long-term maintainability across the engineering team.
• Designed and built the architecture for a fleet of highly automated, lead-generating microservices. This scalable system rapidly became the primary source of inbound leads for the product, requiring minimal developer intervention for deployment and A/B testing.

Stack: Vue 2/3, Nuxt 2/3, JavaScript/TypeScript, Vuex/Pinia, NodeJS, SSR, SEO/GEO Optimization, FSD Architecture, Django/Python (API Integration), PostgreSQL.

January 2020 — June 2023

One of the largest freelance marketplaces in the CIS, serving thousands of clients and professionals.

• Successfully delivered over 20 diverse Fullstack projects, specializing in responsive applications built with Vue.js and Nuxt, ensuring high standards for UX, accessibility, and architectural robustness.
• Performed extensive codebase refactoring and dependency pruning across various client projects, achieving a reduction in bundle size by up to 40% and consistently improving Google Lighthouse scores to 90+.
• Engineered efficient data fetching strategies by integrating REST and GraphQL APIs. Implemented aggressive client-side caching (e.g., via Vuex/Pinia), resulting in page load time improvements of up to 50%.
• Conducted data-driven UI/UX improvements based on A/B testing for e-commerce clients, directly resulting in an average 30% increase in conversion rates.
• Designed and implemented multiple reusable component libraries (Design Systems) using composition principles, accelerating future feature development for clients by an estimated 25%.

Stack: Vue.js, Nuxt.js, Vuex/Pinia, JavaScript/TypeScript, REST/GraphQL, Figma, A/B Testing, PostgreSQL/MongoDB.