Pavel Masharov
Python Developer
6+ years exp
10+ projects
C#
Python
SQL
Data Pipeline & Feature Engineering: 4/10
Active 18 days ago
+7 (982) 6604550 Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
A practical middle-level LLM engineer who designs and implements end-to-end agentic RAG systems and media-processing pipelines. The strongest proven skill is building agent workflows and retrieval pipelines, evidenced by the LangGraph-based agent graphs (agent/graph.py, agent-researcher/graph/workflow.py) together with the retrieval/ingestion code (retrieval/retriever.py, ingestion/loaders.py, ingestion/splitters.py). There is little evidence of custom model training, experiment tracking or production-grade MLOps (CI/CD, monitoring, model/version management) in the public code.
Technical skills
C#• Senior • 6y+ • 10+ projects
Python• Middle
SQL• Middle
Python
Aiohttp
aiogram
Asyncio
FastAPI
Databases
PostgreSQL
Databases
MySQL• 4y+ • 5+ projects
FAISS
Chroma• 4 projects
AI/ML
Fine-tuning
Prompt Engineering
Transformers
OpenAI SDK
Streamlit
Pandas
LangChain
LLM
NumPy
PyTorch
RAG
Scikit-learn
Sentence-Transformers
Whisper
NLP
DevOps
Cloudflare
Git
Rest API
WebSockets
Azure DevOps• 3 projects
Mobile
Firebase
Android SDK• 4y+ • 4 projects
iOS SDK• 4y+ • 4 projects
Timeline
IT-Krypto Team Lead
•
Lead
IT-Krypto
•
Full-Time
Led development of AI/ML integration projects, building LLM-based services for automated information analysis. Designed and built an enterprise CRM with LLM-driven analysis and meeting summaries, and implemented async Telegram bots for crypto fintech integrations using exchange APIs. Also architected and delivered VPN/proxy infrastructure, creating a FastAPI backend to manage WireGuard and Shadowsocks with automation and billing APIs.
Pythonsince 2024
FastAPIsince 2024
LangChainsince 2024
aiogramsince 2024
Asynciosince 2024
Rest APIsince 2024
WebSocketssince 2024
FAISSsince 2024
Sentence-Transformerssince 2024
Scikit-learnsince 2024
PyTorchsince 2024
NumPysince 2024
Whispersince 2024
LLMsince 2024
RAGsince 2024
Team Lead
•
Lead
IT-Krypto
•
Full-Time
Led development of AI/ML-driven services for analyzing information from crypto exchanges, crypto services, and external sources such as social channels and websites using LLMs. Built an enterprise CRM with LLM-based analysis of operational metrics and automated meeting summaries. Developed async Telegram bots for fintech/crypto integrations with exchange APIs and high-frequency data processing for trading signals and order placement. Designed and implemented a FastAPI-based VPN/proxy backend to manage server/client infrastructure and provide automation for deployment and billing.
Python
FastAPI
LangChain
RAG
Whisper
PyTorch
Sentence-Transformers
Scikit-learn
FAISS
NumPy
Asyncio
aiogram
Rest API
WebSockets
SQL
LLM
Unity Developer
•
Middle
Bimi Boo Kids
•
Full-Time
Developed and maintained Unity applications by implementing new features, SDK/plugin changes, and ongoing support for existing releases. Conducted code testing and debugging, validated business logic, and created technical documentation. Integrated Firebase analytics into an application and later wrote integration documentation for reuse in other apps. Delivered localization to 42 languages using I2 Localization and built a Cloudflare-backed content delivery service to reduce app size.
Unity Developer
•
Middle
TurboGames
•
Full-Time
Built new Unity applications and implemented product modifications such as features and SDK/plugin updates while maintaining existing functionality. Wrote quality code according to provided specifications and performed code testing and debugging. Reviewed logic, fixed issues, and produced technical documentation to support development and maintenance.
Ural Law Institute of the Ministry of Internal Affairs of Russia
Bachelor's Degree •
Law enforcement; Jurisprudence
Middle AI/ML Engineer
Confidence: Medium LLM Engineer
A practical middle-level LLM engineer who designs and implements end-to-end agentic RAG systems and media-processing pipelines. The strongest proven skill is building agent workflows and retrieval pipelines, evidenced by the LangGraph-based agent graphs (agent/graph.py, agent-researcher/graph/workflow.py) together with the retrieval/ingestion code (retrieval/retriever.py, ingestion/loaders.py, ingestion/splitters.py). There is little evidence of custom model training, experiment tracking or production-grade MLOps (CI/CD, monitoring, model/version management) in the public code.
Model Architecture & Training
1/10
How well models are designed and trained
Minimal model engineering - code initializes and calls hosted LLMs (ChatOpenAI) but contains no training, custom architectures, or tuning pipelines.
Evidence
video_cutter_ai/src/video_processor/ai_analyzer.py: _initialize_llm (ChatOpenAI initialization)
agent-researcher/src/agents/researcher.py: llm = ChatOpenAI(model=Config.LLM_MODEL, temperature=0.2)
Data Pipeline & Feature Engineering
4/10
How data is prepared for models
Clear, modular data ingestion and retrieval pipeline with multi-format loaders, structured chunking and a retriever abstraction for vector search.
Evidence
rag-agent/src/ingestion/loaders.py: load_all_documents, load_markdown_files, load_pdf_files
rag-agent/src/ingestion/splitters.py: split_by_markdown_headers, split_simple
rag-agent/src/retrieval/retriever.py: DocumentRetriever.search and get_available_sources
Experimentation & Evaluation
2/10
How results are measured and tested
Basic evaluation/metrics and logging exist for pipeline steps, but no formal experiment tracking, validation suites, or ablation infrastructure.
Evidence
video_cutter_ai/src/video_processor/segment_analyzer.py: analysis_metrics (coverage/average_segment_duration) and logging
agent-researcher/src/graph/workflow.py: structured workflow orchestration used for multi-step agent runs
MLOps & Deployment
3/10
How models are shipped to production
Practical MLOps elements for ingestion, persistence and lightweight serving (Streamlit UI, vector store creation, ingestion runner), but no CI/CD, model versioning or monitoring.
Evidence
rag-agent/src/ingestion/pipeline.py and run_ingestion.py: run_ingestion_pipeline entrypoints
rag-agent/src/ui/streamlit_app.py and agent-researcher/src/ui/streamlit_app.py: Streamlit UIs
rag-agent/src/ingestion/embeddings.py: create_vector_store / load_vector_store (Chroma integration)
Computational Efficiency
1/10
How efficiently computing resources are used
Little to no work targeting computational efficiency - no quantization, GPU optimization, batching strategies or profiling artifacts.
Evidence
video_cutter_ai/src/video_processor/video_processor.py: end-to-end pipeline orchestration (no GPU/efficiency instrumentation)
video_cutter_ai/src/video_processor/video_cutter.py: straightforward moviepy usage without performance tuning
Research Depth & Innovation
3/10
Depth of research and new ideas
Applied engineering and pragmatic design (agent graphs, semantic segmentation heuristics, robust prompt handling) but no novel research contributions or reproduced paper results.
Evidence
agent-researcher/src/graph/workflow.py: multi-agent graph orchestration with typed StateGraph
video_cutter_ai/src/video_processor/segment_analyzer.py: non-trivial segmentation/splitting logic and analysis metrics
video_cutter_ai/src/video_processor/ai_analyzer.py: structured prompt/response parsing and validation
Verified artifacts
Expertise
AI Agents & Agentic Workflows• Middle
RAG• Middle
Audio & Speech Processing• Middle
Industries
Artificial Intelligence• Middle
Media & Entertainment• Middle
Technologies
SQL• Middle
PostgreSQL
Rest API
LangChain
FAISS
Sentence-Transformers
FastAPI
WebSockets
Fine-tuning
Scikit-learn
Prompt Engineering
NLP
OpenAI SDK
Transformers
Pandas
NumPy
Git
PyTorch
Cloudflare
LLM
RAG
Asyncio
Whisper
Streamlit
Aiohttp
aiogram
Recommendations
- Develop and own RAG/agent integrations - build production RAG agents and orchestration (retrieval tuning, tool interfaces, guardrails).
- Implement audio/video ML pipelines for media products - extend the existing video pipeline with batching, profiling and monitoring for scale.
- Add experiment tracking and CI for model-driven components - integrate W&B/MLflow and add unit/integration tests for pipelines and agent behaviors.
- Harden MLOps and deployment - add model/versioning, inference benchmarks, latency/cost budgets and simple drift/health checks.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Backend Developer
Confidence: Medium Generalist
A pragmatic backend-focused generalist at a mid-level who can design and implement end-to-end AI pipelines in Python. The strongest proven skill is building LLM-driven pipelines and retrieval-augmented flows with concrete artifacts such as the DocumentRetriever (src/retrieval/retriever.py) and LangGraph workflows (agent-researcher/src/graph/workflow.py). The public code lacks evidence of production-grade concerns like distributed resilience, formal API versioning/idempotency, migration histories, and automated testing.
API Design
2/10
How well APIs are designed
API-level design exists (tool wrappers and LLM invocation) but lacks formal versioning, idempotency, consistent error contract, or pagination/limit strategies; evidence shows pragmatic tool functions rather than a designed API surface for external clients.
Evidence
rag-agent/src/agent/tools.py: tool definitions (search_documentation, list_documentation_sources, search_in_specific_document)
agent-researcher/src/agents/researcher.py: LLM invocation and tool usage pattern
Data Layer & Database
3/10
Working with databases
Uses a vector DB (Chroma) via a vector-store wrapper and maintains metadata; there is basic error handling and metadata extraction, but no migration history, advanced indexing or tuned query/SQL work evident.
Evidence
rag-agent/src/retrieval/retriever.py: DocumentRetriever._init_vector_store, search(), get_available_sources()
rag-agent/src/retrieval/vector_store.py: get_vector_store (vector store initialization pattern)
Scalability & Performance
2/10
Handling load and speed
Pipeline-level batching and chunking are used (text splitters, batch video processing) but there are no explicit caching invalidation strategies, queue-based decoupling, rate-limiting, or measured performance optimizations.
Evidence
rag-agent/src/ingestion/splitters.py: chunking strategy with Config.CHUNK_SIZE and RecursiveCharacterTextSplitter
video_cutter_ai/src/video_processor/video_processor.py: process_all_videos loop and end-to-end pipeline orchestration
System Architecture
4/10
Overall system structure
Clear modular decomposition and purposeful pipeline design (agents, ingestion, retrieval, UI, utils) and LangGraph workflows indicate deliberate system structuring, but there is no evidence of service-level contracts, multi-service deployments or distributed resilience design.
Evidence
agent-researcher/src/graph/workflow.py: StateGraph-based workflow and nodes (plan, research, write)
video_cutter_ai/src/video_processor/video_processor.py: composition of AudioExtractor, Transcriber, SegmentAnalyzer, AIAnalyzer, VideoCutter into a single pipeline class
Security & Auth
2/10
Protecting data and access
Some basic secrets checks and input size guards exist, and environment-based API key checks are present, but there is limited evidence of a comprehensive authn/authz model, input validation at all boundaries, or dependency security/audit processes.
Evidence
video_cutter_ai/src/video_processor/ai_analyzer.py: _initialize_llm checks Config.OPENAI_API_KEY and raises/logs on missing key
rag-agent/src/agent/mcp_tools.py: read_file enforces max_lines and file-size checks before returning content
Reliability & Observability
3/10
Stability and monitoring
Extensive pragmatic logging and try/except fallbacks are present across loaders, retrievers and processors; structured observability (metrics, correlation ids), retries with backoff, and circuit-breakers are not present.
Evidence
rag-agent/src/ingestion/loaders.py: try/except per file loader with logger.info/error and aggregated loading stats
video_cutter_ai/src/video_processor/segment_analyzer.py: robust logging, try/except in analysis flow, save_analysis and analysis metrics computation
Expertise
Backend AI & LLM• Middle
Databases & Vector Storage• Middle
Python• Middle
Industries
Media & Entertainment• Middle
Software• Middle
Technologies
Databases
Python• Middle
Recommendations
- Develop LLM-backed RAG agents and internal knowledge search services (ingestion, chunking, vectorization, retriever integration).
- Build end-to-end AI media processing pipelines that include transcription, semantic segmentation and automated clipping (use the video_processor pipeline).
- Implement production hardening tasks: add structured metrics, retries with exponential backoff, rate limiting, and integration tests for the vector-store and LLM interactions.
- Improve API contracts and add idempotency/versioning and migration history when evolving the vector store schema
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: