Maxim Petrov

Builds and operates production LLM automation systems end to end on a self-hosted stack, from client interviews and architecture through deployment and ongoing support. Leads agentic, multi-step pipelines with human-in-the-loop controls and strict input/output contracts to manage token cost and reliability.

Built a Telegram-based intake bot that performs multi-stage human-in-the-loop processing (voice/text preferences, style detection, confirmations, and iterative reformulation) and generates room visualizations. Used Supabase for centralized storage of user data and session artifacts, with Whisper and OpenAI image generation in the pipeline.

Developed an inbound request and document parsing module that classifies requests, performs structured data extraction, and returns JSON output as a contract to a client system. Used PostgreSQL-backed issue tracking to capture classification errors for prompt refinement and testability.

Implemented an on-site photo-report analysis pipeline where crews submit voice notes and photos, and managers receive automatic reports and escalations for critical findings. Routed mixed content types in a single workflow and formatted outputs for downstream task tracking.

Built a procurement forecasting and stock control workflow that recognizes staff voice notes, records stock updates in PostgreSQL, and flags early shortage-risk signals. Prototyped supplier sourcing using Firecrawl and ScrapeGraph AI and exposed results through Notion.

Created a recording-based meeting analysis pipeline that pulls files from Google Drive, runs async AssemblyAI diarization, and sends results via webhook with speaker mapping, summaries, and negotiation assessment. Implemented a low-cost speaker mapping approach with reroll and manual labeling to reduce ID-error cost.

Created a scoping and pricing assistant for Digital Cinema Package services by capturing requirements through a structured, recorded, diarized interview and encoding business logic to remain editable without code rewrites. Implemented flexible agent memory keyed per request and a human-in-the-loop issue tracker to improve prompt quality.

Built a multi-step, n8n-based agentic pipeline to collect YouTube review sources, extract transcripts, summarize stepwise, and produce a final reviewed output with human-in-the-loop source selection. Added AI sales voiceover and reduced review preparation time significantly for car processing workflows.

Grade: Honors diploma

Bachelor's program in Public and Municipal Administration, completed with an honors diploma. Combined formal study with self-directed learning through production projects.