Dmitriy Denisov
Python Developer
5+ years exp
SQL
TypeScript
JavaScript
PHP
Node JS
Python
Data Pipeline & Feature Engineering: 5/10
MLOps & Deployment: 5/10
Computational Efficiency: 4/10
Active 18 days ago
+7 (999) 5348895 Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
LLM engineer (senior-level) specializing in building production-oriented LLM orchestration and document-intelligence pipelines. The strongest proven skill is designing robust LLM-to-tool orchestration and function-calling loops as implemented in backend/app/llm_client.py together with the MCP tools in backend/app/mcp_server.py. There is little to no evidence of custom model training, experiment tracking, formal unit tests or performance benchmarking in the public code.
Technical skills
SQL
TypeScript
JavaScript• Middle • 5y+
PHP• Middle • 5y+
Node JS• Junior
Python• Senior
Node JS
Express
Python
Requests
FastAPI
Pydantic
HTTPX
Uvicorn
Asyncio
Databases
PostgreSQL
MySQL• 5y+
Supabase
AI/ML
Chain-of-Thought
Gemini
OpenRouter
Streamlit
DevOps
Git
Frontend
Nuxt.js
Vue.js
Timeline
Python/AI Developer
•
Middle
IP Chernenko S. V.
•
Full-Time
Developed AI tools from scratch to automate business processes, including system architecture, LLM integration, ETL pipelines, prompt engineering, and deployment. Built an analytics dashboard that processes data from multiple sources and updates reports by period. Implemented a local-LLM-based AI assistant and created a PDF document auto-checking tool using Gemini API with chunking and retry logic. Delivered an admin panel with role-based access control.
Python
Streamlit
vLLM
Qwen
Gemini
OpenRouter
Chain-of-Thought
Fullstack Developer (Vue/Nuxt)
•
Middle
IP Antoshkin K. V.
•
Full-Time
Built a product catalog for a furniture company, covering architecture, frontend and backend development, performance optimization, and production deployment. Implemented advanced search with fuzzy matching and autocomplete, plus custom slider and gallery components. Added UI features such as skeleton loading, filtering, sorting, and SPA routing. Configured Nuxt.js SSR to improve SEO and Core Web Vitals while keeping the UI component-free using native solutions.
Vue.js
Nuxt.js
Node JS
JavaScript
Supabase
Rest API
Frontend/Fullstack Developer
•
Middle
Freelance
•
Freelance
Delivered end-to-end commercial web projects for multiple industries, handling design-to-implementation layout, backend development, integrations, and production release. Developed interactive website features such as an SVG floor plan with hover highlighting and PDF plan exports. Implemented a credit calculator using formula-based computations and built e-commerce functionality with email and OAuth-based authentication plus an admin area. Created responsive, pixel-perfect HTML/CSS layouts and integrated payment services such as Stripe and other external APIs.
JavaScriptsince 2021
PHP
MySQL
Rest APIsince 2021
Senior AI/ML Engineer
Confidence: High LLM Engineer
LLM engineer (senior-level) specializing in building production-oriented LLM orchestration and document-intelligence pipelines. The strongest proven skill is designing robust LLM-to-tool orchestration and function-calling loops as implemented in backend/app/llm_client.py together with the MCP tools in backend/app/mcp_server.py. There is little to no evidence of custom model training, experiment tracking, formal unit tests or performance benchmarking in the public code.
Model Architecture & Training
How well models are designed and trained
Not evidenced in public code
Data Pipeline & Feature Engineering
5/10
How data is prepared for models
Solid data ingestion and cleaning code for DOCX/PDF extraction, chunking and Excel parsing with explicit edge-case handling.
Evidence
tz-drawing-analyzer/core.py: pdf_to_text (page/block extraction and ordering)
tz-drawing-analyzer/core.py: _extract_chunked (chunking and merge logic for large PDFs)
biocad/backend/app/excel_io.py: parse_excel and build_excel (openpyxl-based Excel parsing and export)
Experimentation & Evaluation
2/10
How results are measured and tested
Basic evaluation and reporting pipeline with summary counters and human-readable reports but no experiment tracking or reproducible ML evaluation framework.
Evidence
tz-drawing-analyzer/core.py: compare_parameters (normalization rules and summary counters)
tz-drawing-analyzer/core.py: generate_report and run.py: print_summary (reporting / summary metrics)
MLOps & Deployment
5/10
How models are shipped to production
Practical service and deployment decisions including FastAPI endpoints, in-process MCP transport, async HTTP usage and explicit API-key handling, but no CI/CD manifests or monitoring/observability integration.
Evidence
biocad/backend/app/main.py: lifespan using create_connected_server_and_client_session to run MCP server in-process
biocad/backend/app/llm_client.py: async _call_openrouter using httpx and _tools_cache to avoid repeated schema fetches
Computational Efficiency
4/10
How efficiently computing resources are used
Reasonable efficiency choices such as chunking, caching and exponential backoff; lacks GPU/quantization work or measured profiling results.
Evidence
tz-drawing-analyzer/core.py: _extract_chunked chunking strategy for large documents
biocad/backend/app/llm_client.py: _tools_cache and MAX_TOOL_ITERATIONS; core.call_llm uses retry with exponential backoff
Research Depth & Innovation
3/10
Depth of research and new ideas
Thoughtful prompt engineering, normalization and verification rules and a multi-step extraction-verification-comparison pipeline, but no novel research, custom layers or reproduced paper implementations.
Evidence
tz-drawing-analyzer/core.py: compare_parameters has detailed normalization rules and classification logic
biocad/backend/app/llm_client.py: LLM-to-tool orchestration loop implementing function/tool calling with iteration limits
Expertise
Document Intelligence & OCR• Senior
AI Agents & Agentic Workflows• Senior
Industries
Software• Middle
Technologies
SQL
MySQL• 5y+
PostgreSQL
Supabase
Chain-of-Thought
Git
Gemini
Streamlit
OpenRouter
Recommendations
- Develop LLM-powered agentic services and tool integrations - function-calling orchestrators, tool schemas and robust tool-result handling.
- Build document-intelligence pipelines for engineering and construction documents - PDF/DOCX extraction, chunking, verification and normalization.
- Productionize FastAPI services with CI/CD, automated tests and observability (metrics, logs, error budgets).
- Add experiment tracking and formal evaluation for extraction accuracy and prompt variants (W&B, MLflow or similar) and perform performance profiling.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Senior Backend Developer
Confidence: High API Engineer
Senior backend engineer specializing in API-driven LLM orchestration and domain-specific automation for plan scheduling and technical-document analysis. The strongest proven skill is LLM-tool orchestration and domain integration as shown by llm_client.run_chat_turn together with the MCP toolset in mcp_server.py. Public code lacks persistent storage, formal testing, production observability and deployment hardening evidence.
API Design
5/10
How well APIs are designed
API design shows deliberate endpoint contracts, response models and error handling but lacks formal versioning, pagination, or idempotency machinery and is largely a focused domain API rather than a general REST platform.
Data Layer & Database
2/10
Working with databases
No external database or migration history; state is in-memory with explicit domain logic, so data-layer concerns are limited to in-memory consistency rather than transactional persistence or tuned queries.
Scalability & Performance
4/10
Handling load and speed
Performance and scaling trade-offs are visible (in-process MCP transport to avoid IPC, chunking large documents, timeouts and retries), but there is no evidence of capacity planning, caching/invalidation, connection pooling tuning or rate limiting.
Evidence
drk-supersonic/biocad/backend/app/main.py: create_connected_server_and_client_session in lifespan - in-memory MCP transport to avoid separate process IPC
drk-supersonic/tz-drawing-analyzer-v-3/core.py: _extract_chunked implements chunking/merging for large PDFs
drk-supersonic/biocad/backend/app/llm_client.py: _call_openrouter uses httpx.AsyncClient with timeout=60 and run_chat_turn limits iterations (MAX_TOOL_ITERATIONS)
System Architecture
5/10
Overall system structure
Clear separation of concerns and deliberate architecture choices (FastAPI host + MCP tool server in-process, LLM orchestration layer, CLI/Streamlit UIs), showing thoughtful decomposition for the product; lacks service-to-service contracts or multi-process deployment docs but the design is purposeful.
Evidence
drk-supersonic/biocad/backend/app/mcp_server.py: FastMCP tool definitions expose domain actions (list_tasks, add_task, move_task, etc.)
drk-supersonic/biocad/backend/app/main.py: FastAPI + lifecycle creating connected MCP session; REST endpoints delegate to MCP tools
drk-supersonic/tz-drawing-analyzer-v-3: separation between core analysis (core.py), CLI (run.py) and Streamlit UI (app.py)
Security & Auth
4/10
Protecting data and access
Security-aware choices are present (per-request API key use, not persisted, input validation on Excel parsing, pydantic models), but there is limited evidence of hardened auth/authorization, dependency auditing, or defenses against LLM-specific threats beyond pragmatic checks.
Evidence
drk-supersonic/biocad/backend/app/llm_client.py: _call_openrouter raises on missing/empty api_key and top-level comment states API key is not persisted
drk-supersonic/biocad/backend/app/main.py: chat endpoint catches RuntimeError for missing key and returns HTTP 400; upload_excel validates columns in excel_io.parse_excel
drk-supersonic/biocad/backend/app/models.py: use of Pydantic models (response_model in endpoints) implied for input validation
Reliability & Observability
4/10
Stability and monitoring
Reliability patterns are used (HTTP timeouts, retries with exponential backoff, iteration limits, structured debug output), but there is no evidence of metrics/alerts, distributed tracing or circuit breakers and graceful shutdown beyond the FastAPI lifespan context.
Evidence
drk-supersonic/tz-drawing-analyzer-v-3/core.py: call_llm implements retries with exponential backoff and timeouts on requests.post
drk-supersonic/biocad/backend/app/llm_client.py: MAX_TOOL_ITERATIONS and try/except around tool calling to avoid infinite loops; httpx.AsyncClient with timeout
drk-supersonic/biocad/backend/app/main.py: lifespan contextmanager to create and clean up MCP session and result._debug returned for debugging
Expertise
Backend AI & LLM• Senior
Microservices & API Architecture• Middle
Python• Middle
Industries
Science & Engineering• Middle
Technologies
PHP• Middle • 5y+
FastAPI
Asyncio
Pydantic
HTTPX
Uvicorn
Recommendations
- Develop LLM-driven backend orchestration services and internal tools that integrate model tool-calls with deterministic domain logic (e.g., scheduling, document verification).
- Build mid-sized API services that require custom business logic and in-memory coordination (task planners, workflow engines) where the in-process MCP pattern and tool-calling loop are useful.
- Prototype domain-specific document analysis pipelines combining Python extraction, chunking and LLM post-processing (construction/engineering QA workflows).
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle DevOps Engineer
Confidence: High Generalist
Middle-level engineer specializing in LLM-driven document analysis pipelines for construction drawing compliance checks. The strongest proven skill is designing and orchestrating an LLM-based extraction/verification/comparison pipeline implemented in core.py (discover_parameters, extract_parameters, _extract_chunked, verify_extracted, compare_parameters). Public code does not evidence CI/CD, IaC, container orchestration, observability-as-code, formal testing, or cloud architecture artifacts.
CI/CD Pipelines
Automated build and deploy
Not evidenced in public code
Infrastructure as Code
Managing servers with code
Not evidenced in public code
Containerization & Orchestration
Working with containers
Not evidenced in public code
Observability & Monitoring
Watching system health
Not evidenced in public code
Reliability & Incident Response
2/10
Keeping systems up
Basic operational resilience and incident-avoidance patterns exist in application code: API timeouts, retry with exponential backoff, chunking to avoid payload limits, input validation and graceful error messages, but there are no runbooks, deploy strategies, chaos or postmortems.
Evidence
tz-drawing-analyzer-v-3/core.py: call_llm implements requests.post(..., timeout=300) and retry with exponential backoff and raises structured RuntimeError on repeated failures
tz-drawing-analyzer-v-3/core.py: _extract_chunked splits large text and merges chunked results to avoid single huge requests
tz-drawing-analyzer-v-3/run.py: ask_api_key and find_files perform input validation and user-friendly error exits; print_progress provides progress feedback
Cloud & Cost Optimization
Smart use of the cloud
Not evidenced in public code
Technologies
Python• Senior
Requests
Recommendations
- Extend this work into a reproducible deployment pipeline (CI/CD) that builds, tests, and deploys the app with environment separation and secrets stored in a vault or external secret manager.
- Add unit and integration tests (mocking the LLM calls) and infra tests (terratest/conftest) so extraction and normalization logic are verifiable and regression-safe.
- Containerize the application with a multi-stage Dockerfile and add a simple Kubernetes or cloud run deployment manifest with resource requests/limits and liveness/readiness probes.
- Add observability: structured logging, request/latency metrics (Prometheus) and basic alerting/SLOs for API failures and latency to detect model/service regressions.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: