Yaroslav Malovichko
Python
Data Wrangling & Cleaning: 7/10
Business Insight & Impact: 6/10
Reproducibility & Notebook Hygiene: 5/10
Active 18 days ago
Invite to interview
Message
Download CVCV
Overview
Technical skills
Roles
Overview
A pragmatic data engineer at a lower-senior level who builds end-to-end domain-specific ETL and RAG pipelines. The strongest proven skill is robust, domain-aware data ingestion and diagnostic logic, evidenced by the TailingsReport parser and the diagnostics -> LossChannel aggregation used to drive hypothesis generation. What is not evidenced is predictive-model development, large-scale distributed systems design, or full MLOps/CI at production scale.
Technical skills
Python• Senior
Python
FastAPI
Uvicorn
Pydantic
AI/ML
Streamlit
Pandas
LangChain
Anthropic SDK
OpenAI SDK
huggingface_hub
Senior Data Scientist
Confidence: Medium Data Engineer
A pragmatic data engineer at a lower-senior level who builds end-to-end domain-specific ETL and RAG pipelines. The strongest proven skill is robust, domain-aware data ingestion and diagnostic logic, evidenced by the TailingsReport parser and the diagnostics -> LossChannel aggregation used to drive hypothesis generation. What is not evidenced is predictive-model development, large-scale distributed systems design, or full MLOps/CI at production scale.
Statistical Rigor
2/10
Correct use of statistics
Minimal statistical rigor - simple arithmetic cross-checks and data warnings are present, but no formal hypothesis testing, uncertainty quantification or multiple-comparison controls.
Evidence
hypothesis_factory/src/diagnostics.py: crosscheck logic in diagnose() that compares calculated vs file totals and emits warnings
hypothesis_factory/src/tailings_parser.py: _num() robust numeric parsing and rep.warnings for broken/absent data
Data Wrangling & Cleaning
7/10
Preparing and cleaning data
Strong, pragmatic data wrangling for messy Excel reports - robust parsing, normalization, error-tolerant numeric conversion and explicit warnings; good domain-aware data model.
Evidence
hypothesis_factory/src/tailings_parser.py: parse_tailings_xlsx(), _norm_class(), _num() and data model classes (TailingsReport, TailingsStream, SizeClassBlock)
hypothesis_factory/src/diagnostics.py: construction of LossChannel and Diagnosis, recoverable_t aggregation logic
Exploratory Analysis & Visualization
4/10
Exploring and visualizing data
Basic exploratory outputs and storytelling via Streamlit and textual summaries, but EDA is targeted to domain reporting rather than statistical exploration or visual analytic depth.
Evidence
hypothesis_factory/src/app.py: Streamlit UI presenting metrics, bar_chart and dataframes
hypothesis_factory/src/diagnostics.py: _summary_ru() and render_report() producing human-readable interpretation
Predictive Modeling
2/10
Building models that predict
No predictive-model development; the work focuses on retrieval, prompt engineering and LLM orchestration with a simple scoring formula rather than model training or evaluation pipelines.
Evidence
hypothesis_factory/src/hypothesis_engine.py: generate() implements retrieval + LLM + JSON parsing and a simple weighted scoring function
Degen14/Rag/app/vectorstore.py and app/rag_chain.py: RAG pipeline and embeddings retrieval usage (no model training code)
Business Insight & Impact
6/10
Turning analysis into business value
Clear business/engineering alignment - domain rules mapping to causes and concrete corrective directions, and prompts that require measurable verification plans and effect estimates.
Evidence
hypothesis_factory/src/diagnostics.py: RULES mapping mineral/coarseness to causes and concrete directions
hypothesis_factory/src/prompts.py: SYSTEM_PROMPT requires measurable verification_plan, success_criteria and expected_effect_t
Reproducibility & Notebook Hygiene
5/10
Clean, repeatable analysis
Moderate reproducibility: pinned requirements, small test coverage and CLI/Streamlit entrypoints exist, but no full CI/CD, data-versioning or seeded experiment pipelines shown.
Expertise
Big Data• Middle
Analytics• Middle
Industries
Data & Analytics• Middle
Natural Resources• Middle
Technologies
Python• Senior
LangChain
FastAPI
Anthropic SDK
OpenAI SDK
huggingface_hub
Pandas
Pydantic
Recommendations
- Lead development of RAG-enabled analytic tools and prototypes that turn messy domain reports into structured facts (ingest/parsers + retriever + prompt templates).
- Build domain-specific ETL pipelines for industrial data (Excel/PDF/scan OCR) with validation and monitoring based on the Tailings parser patterns.
- Implement LLM orchestration for hypothesis generation and feedback loops, including vector search and strict JSON parsing workflows.
- Own productionization of small ML/LLM services (FastAPI + vectorstore) and API endpoints, with focus on test coverage and reproducible deployment.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle AI/ML Engineer
Confidence: Medium LLM Engineer
A mid-level LLM-focused engineer who builds retrieval-augmented generation systems, prompt scaffolding and domain-specific ingestion pipelines. The strongest proven skill is domain-aware data ingestion and engineering, evidenced by the robust tailings parser and diagnostics pipeline (src/tailings_parser.py and src/diagnostics.py) that handle messy Excel reports and produce structured fact bases for RAG. There is little to no evidence of custom model training, measured efficiency work (quantization/profiling), distributed training or mature MLOps/experiment-tracking in the public code.
Model Architecture & Training
1/10
How well models are designed and trained
No custom model architecture or training pipelines are present; work is limited to prompt engineering and thin LLM client wrappers rather than model development.
Evidence
hypothesis_factory/src/prompts.py: SYSTEM_PROMPT and build_user_prompt / build_fact_base (prompt engineering)
hypothesis_factory/src/llm_client.py: OpenAICompatClient, AnthropicClient and MockLLM classes (LLM client wrappers)
Data Pipeline & Feature Engineering
4/10
How data is prepared for models
Solid domain data engineering: robust, defensive parsing of messy Excel reports, chunking/ingest pipelines and OCR integration for building a knowledge base.
Evidence
hypothesis_factory/src/tailings_parser.py: parse_tailings_xlsx implementation with normalization and robust numeric parsing
hypothesis_factory/src/ingest.py and hypothesis_factory/src/ocr_scan.py: chunking, page extraction and parallel OCR pipeline (ProcessPoolExecutor)
Experimentation & Evaluation
3/10
How results are measured and tested
Basic evaluation and validation practices: LLM output parsing/validation and a feedback loop are implemented, plus lightweight API tests, but no formal experiment tracking or systematic A/B evaluation.
Evidence
hypothesis_factory/src/prompts.py: extract_json for robust parsing and validation of LLM JSON outputs
Degen14/Rag/tests/test_api.py: FastAPI endpoint tests for health/ingest/query/agent
MLOps & Deployment
3/10
How models are shipped to production
Contains service components and deployment-ready pieces (FastAPI, Streamlit, cached retriever and LangChain/Chroma integrations) suitable for small-scale deployment but lacks mature MLOps features like model versioning, CI-driven pipelines, monitoring or drift detection.
Evidence
Degen14/Rag/app/main.py: FastAPI endpoints for ingest/query/agent
hypothesis_factory/src/app.py: Streamlit application with st.cache_resource and export functionality
Degen14/Rag/app/rag_chain.py: get_llm() using ChatOllama integration
Computational Efficiency
2/10
How efficiently computing resources are used
Some attention to performance (parallel OCR, optional GPU embeddings mentioned), but no measured profiling, quantization, batching optimizations or documented GPU/latency tradeoffs.
Evidence
hypothesis_factory/src/ocr_scan.py: use of ProcessPoolExecutor for parallel OCR
hypothesis_factory/requirements.txt: optional comments about sentence-transformers (GPU) and vllm (local LLM)
Research Depth & Innovation
2/10
Depth of research and new ideas
Applied domain heuristics and a novelty check indicate practical innovation, but there is no evidence of original research, new algorithms or paper-level reproductions.
Evidence
hypothesis_factory/src/diagnostics.py: domain expert RULES and match_rule logic for interpreting channels of loss
hypothesis_factory/src/hypothesis_engine.py: novelty checking using Jaccard overlap and feedback integration
Expertise
LLM• Middle
RAG• Middle
AI Agents & Agentic Workflows• Middle
Industries
Natural Resources• Middle
Technologies
Streamlit
Uvicorn
Recommendations
- Lead development of RAG-backed assistants and domain-specific LLM workflows (ingest, chunking, retrieval, prompt scaffolding)
- Build productionization around the existing services - add CI/CD, model/version gating, monitoring and drift detection for the RAG/LLM pipeline
- Expand validation and experiment tracking (W&B/MLflow) and add quantitative profiling to justify any GPU/embedding choices
- Harden security of agent/tooling code (avoid eval, sanitize subprocess usage) before deploying agent features in production
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: