Natalia Antropova
AI Engineer
SQL
C#
Node JS
Swift
JavaScript
Objective-C
C
Java
Python
TypeScript
Experimentation & Evaluation: 6/10
Data Pipeline & Feature Engineering: 5/10
MLOps & Deployment: 5/10
Active 22 days ago
+7 (999) 7159189 Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
Senior AI Product Engineer building and shipping AI-first products end to end—from product discovery and UX to full-stack architecture, LLM/RAG pipelines, evaluation, monetization, and production operations. I have launched revenue-generating SaaS, Telegram, and native macOS products and delivered client LLM platforms, browser extensions, and media systems. My background combines an MSc in Machine Learning Engineering with 6+ years in enterprise QA, giving me a strong focus on measurable quality, privacy, reliability, failure recovery, and production readiness.
Technical skills
SQL
C#
Node JS• Middle
Swift
JavaScript
Objective-C
C
Java• Middle • 3y+
Python• Senior
TypeScript• Middle
Node JS
BullMQ
Prisma
C
FFmpeg• 3y+
Python
FastAPI
SQLAlchemy
HTTPX
Databases
PostgreSQL
Redis
MySQL• 3y+
AI/ML
AI Agents
Claude
Cline
Copilot
Cursor
Kling
LLM
Mistral
Mistral SDK
Prompt Engineering
RAG
Windsurf
TF-Keras
Streamlit
Embeddings
Frontend
hls.js
Next.js
React Three Fiber
React.js
Tailwind CSS
Three.JS
PostCSS
Zod
Vite
DevOps
CI/CD
Docker
Docker Compose
Git
Rest API
Vector
Design
Figma
Mobile
AVFoundation
MVP
PWA
StoreKit
SwiftUI
TestFlight
QA
Playwright
Postman• 5y+
Swagger• 5y+
JMeter• 3y+
Selenium• 3y+
Timeline
Senior QA Engineer
•
Senior
BaccaSoft
•
Full-Time
Performs manual and automation web/API testing for enterprise systems, including BPM-based workflows with authorization and cryptography-related features. Designs functional, integration, regression, and negative test cases and validates roles, permissions, and edge scenarios. Maintains and updates Selenium UI automation, performs API testing via Postman/Swagger, and analyzes logs. Conducts performance testing using Apache JMeter and documents release acceptance artifacts.
Selenium
Java
Postman
Swagger
MySQL
JMeter
Ural Federal University
Master's Degree •
Machine Learning Engineering
QA Engineer
•
Middle
Extyl
•
Full-Time
Tests corporate web systems (portals, e-commerce, online education) through manual web and API checks, and produces test documentation such as instructions, test cases, and guides. Supports clients with presentations and technical assistance. Works with Postman/Swagger-based API verification and regression for critical scenarios across multiple customer projects, including LMS and document workflow platforms.
Postmansince 2021
Swaggersince 2021
Astrakhan State University
Bachelor's Degree •
Sociology (Social Communications)
Senior AI/ML Engineer
Confidence: High LLM Engineer
A senior-level LLM-focused engineer specializing in retrieval-augmented generation, evaluation and editorial orchestration. The strongest proven skill is building reliable RAG and evaluation tooling, evidenced by the retrieval adapters and llm_eval.evaluate_generation_fixture (src/linked_ai_content_pipeline/rag.py and src/linked_ai_content_pipeline/llm_eval.py). Public code lacks large-scale training experiments, production deployment manifests (Kubernetes/infra IaC) and advanced efficiency work like quantization or GPU profiling.
Model Architecture & Training
3/10
How well models are designed and trained
Custom training loop and loss are present (TF Keras CustomModel) but model work is limited in scale and not accompanied by rigorous reproducible experiment tracking or benchmarking.
Evidence
IMG_Story/models/custom_model.py: CustomModel.train_step and compute_loss
IMG_Story/models/trainer.py: train_model / create_or_load_model
Data Pipeline & Feature Engineering
5/10
How data is prepared for models
Well-structured data collection, normalization and deduplication logic with concurrent collectors and deterministic chunking for retrieval; clear handling of edge cases and validation.
Evidence
src/linked_ai_content_pipeline/normalization.py: normalize_and_deduplicate / canonicalize_url
src/linked_ai_content_pipeline/collectors.py: Collector / FixtureCollector implementations
src/linked_ai_content_pipeline/collection.py: collect_concurrently
Experimentation & Evaluation
6/10
How results are measured and tested
Comprehensive evaluation pipelines for LLM generation and retrieval with stability metrics, p95 latency reporting and well-structured fixtures and tests.
Evidence
src/linked_ai_content_pipeline/llm_eval.py: evaluate_generation_fixture (runs_per_case, p95_latency_ms, stability_rate)
src/linked_ai_content_pipeline/retrieval_evaluation.py: RetrievalEvaluation / evaluate_retrieval
tests/test_llm.py: tests validating provider schema and traceable context
MLOps & Deployment
5/10
How models are shipped to production
Practical MLOps-oriented components - adapters for embedding services, Qdrant REST adapter, owned/borrowed client lifecycles and environment-driven runtime builders for production wiring.
Evidence
src/linked_ai_content_pipeline/rag.py: QdrantVectorStore, OpenAIEmbeddingProvider, retrieval_runtime_from_environment
src/linked_ai_content_pipeline/llm.py: external_draft_provider_from_environment and OpenAIContextDraftProvider (owned client handling)
Computational Efficiency
3/10
How efficiently computing resources are used
Some attention to efficiency (chunking, vector normalization, batching semantics) but no advanced GPU/quantization/throughput engineering or profiling artifacts.
Evidence
src/linked_ai_content_pipeline/rag.py: TextChunker and HashEmbeddingProvider (deterministic embedding and normalization)
src/linked_ai_content_pipeline/llm_eval.py: measurement of latency (latency_ms / p95_latency_ms) in evaluation harness
Research Depth & Innovation
3/10
Depth of research and new ideas
Contains some custom algorithmic choices (deterministic hash embeddings, editorial similarity gating, combined loss in custom TF model) but no claimed SOTA reproduction or deep research artifacts.
Evidence
src/linked_ai_content_pipeline/rag.py: HashEmbeddingProvider implementation
src/linked_ai_content_pipeline/editorial.py: SimilarityGate and editorial revision loop
IMG_Story/models/custom_model.py: non-standard loss combining cross-entropy and embedding-distance term
Expertise
RAG• Senior
AI Agents & Agentic Workflows• Senior
LLM• Senior
Industries
Media & Entertainment• Middle
Internet Services• Middle
Technologies
Python• Senior
SQL
MySQL• 3y+
PostgreSQL
Redis
Copilot
Cursor
Windsurf
Rest API
Claude
Docker Compose
SQLAlchemy
FastAPI
Embeddings
Prompt Engineering
AI Agents
Cline
Mistral SDK
Kling
Mistral
CI/CD
Git
Docker
LLM
RAG
TF-Keras
Streamlit
HTTPX
Vector
Qdrant• mentioned only
Recommendations
- Develop production RAG services and retrieval runtimes (Qdrant + embedding adapters) with monitoring, canary deploys and SLOs.
- Build evaluation- and CI-driven LLM testbeds (stability, latency budgets and reproducible fixtures) for model/provider selection.
- Implement deployment/infra automations (IaC, Kubernetes manifests, Argo/Flux pipelines) and runtime observability for embeddings and agent flows.
- Productize editorial agents and orchestration with cost/latency budgets, retry budgets and drift monitoring.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Backend Developer
Confidence: Medium API Engineer
An experienced backend API engineer at a solid middle level who designs and tests provider adapters, embedding/retrieval logic and editorial pipelines. The strongest proven skill is reliable LLM and retrieval integration with strict input/output validation, demonstrated by OpenAI adapter code, embedding/Qdrant adapters and comprehensive tests (e.g. rag.py, llm adapters and http_client tests). There is limited evidence of production infra like multi-service contracts, DB migrations/transactions, large-scale performance tuning or advanced deployment/config orchestration in public code.
API Design
4/10
How well APIs are designed
Clean, versioned provider adapters and strict schema enforcement for LLM calls are present, but full API contract/versioning and idempotency at HTTP API boundaries are limited to provider adapters rather than public REST surface design.
Evidence
chronica-ai-video-pipeline/src/providers/llm.ts:OpenAIResponsesProvider sets x-prompt-version header and posts a json_schema 'format.strict' body
linked-ai-content-pipeline/src/linked_ai_content_pipeline/llm.py:OpenAIContextDraftProvider validates and enforces strict response shapes (tests exercise validation)
Data Layer & Database
4/10
Working with databases
Vector-store adapters and in-memory vector indexing are implemented with idempotent upsert and careful response validation, but there is no evidence of relational schema migrations, transaction isolation tuning, or complex DB sharding.
Evidence
linked-ai-content-pipeline/src/linked_ai_content_pipeline/rag.py:QdrantVectorStore.upsert uses uuid5(document.id) to implement idempotent upserts
linked-ai-content-pipeline/src/linked_ai_content_pipeline/rag.py:InMemoryVectorStore.query implements exact cosine ranking and stable ordering
linked-ai-content-pipeline/tests/test_rag.py:test_qdrant_adapter_uses_idempotent_upsert_and_query_api
Scalability & Performance
4/10
Handling load and speed
Work shows practical scalability considerations - chunking, embedding batching, async clients and limited concurrency control - but lacks evidence of system-wide caching/invalidation strategies, load testing artifacts, or deployed queue-based decoupling patterns.
Evidence
linked-ai-content-pipeline/src/linked_ai_content_pipeline/http_client.py:RetryPolicy and request_with_retry implement retry/backoff behavior (used by embedding/Qdrant adapters)
linked-ai-content-pipeline/src/linked_ai_content_pipeline/rag.py:OpenAIEmbeddingProvider constructs and reuses httpx.AsyncClient with timeouts
chronica-ai-video-pipeline/tests/pipeline.test.ts:test narration ограничивает конкурентные TTS-вызовы and preserves order (demonstrates concurrency bounding)
System Architecture
4/10
Overall system structure
Modular components (retriever, embedding provider, editorial chain, draft providers) show deliberate boundaries and lifecycle management, but there is limited evidence of multi-service contracts, distributed coordination, or production-grade config/secret orchestration.
Evidence
linked-ai-content-pipeline/src/linked_ai_content_pipeline/editorial.py:EditorialChain composes stages, supports reviser and bounded revision rounds
linked-ai-content-pipeline/src/linked_ai_content_pipeline/rag.py:RetrievalRuntime owns retriever, embeddings and store and performs cleanup on initialization failure
Security & Auth
4/10
Protecting data and access
Input validation, redaction of direct identifiers, environment configuration validation and timeout handling are implemented; full auth lifecycle (token refresh/revocation), dependency audits or SSRF/firewall patterns are not visible.
Evidence
linked-ai-content-pipeline/tests/test_llm.py:test_openai_context_provider_redacts_direct_identifiers_and_bounds_context
chronica-ai-video-pipeline/src/providers/llm.ts:OpenAIResponsesProvider uses AbortController with request timeout and validates/throws on non-OK responses
Reliability & Observability
5/10
Stability and monitoring
Thoughtful reliability patterns are present - retry/backoff with tests, client ownership and clean shutdown, explicit validation and many tests for failure modes - showing stronger operational awareness than typical middle-level projects.
Evidence
linked-ai-content-pipeline/src/linked_ai_content_pipeline/http_client.py:RetryPolicy and request_with_retry (tested in tests/test_http_client.py) implement retries, honor Retry-After and bound connect failures
linked-ai-content-pipeline/src/linked_ai_content_pipeline/rag.py:retrieval_runtime_from_environment closes clients on initialization failure and exposes close() to release network resources
Expertise
Backend AI & LLM• Middle
Databases & Vector Storage• Middle
Python• Middle
Node.js• Middle
Industries
Internet Services• Middle
Media & Entertainment• Middle
Recommendations
- Lead the development of LLM/embedding-backed features - provider adapters, retriever lifecycle and validation harnesses.
- Implement and harden vector-store production integration (idempotency, batch sizing, telemetry) and own Qdrant/embedding deployment runbooks.
- Build end-to-end reliability features: add structured logging/correlation ids, metrics/alerts and documented graceful shutdown for long-running services.
- Expand data-layer expertise by adding migration history, transactional tests or hand-tuned DB queries where needed.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Frontend Developer
Confidence: High Generalist
A senior-level generalist engineer focused on reproducible AI-driven media pipelines and tooling. The strongest proven skill is pipeline orchestration and state discipline - evidenced by the runPipeline manifest/fingerprint/invalidation logic and comprehensive tests in chronica-ai-video-pipeline (src/pipeline/runPipeline.ts and tests/pipeline.test.ts). There is little to no public evidence of original frontend component architecture, accessibility work, or large-scale multi-team front-end system design in the human-authored artifacts.
UI Component Architecture
How interface parts are built
Not evidenced in public code
Responsive & Cross-browser
Works on all screens and browsers
Not evidenced in public code
Performance Optimization
5/10
Speed of the interface
Measured and practical performance work is present - pipeline reuse/fingerprinting to avoid unnecessary work, timeouts and AbortController for external requests, and tests validating concurrency limits and latency measurement.
Evidence
chronica-ai-video-pipeline/src/pipeline/runPipeline.ts: stage fingerprinting and canReuse logic to avoid re-execution
chronica-ai-video-pipeline/src/providers/llm.ts: OpenAIResponsesProvider.request uses AbortController and request timeout handling
chronica-ai-video-pipeline/tests/pipeline.test.ts: tests for concurrent TTS limits and evaluation latency/percentile measurement
Accessibility & Semantics
Usable for everyone
Not evidenced in public code
State Management & Data Flow
6/10
Managing data in the app
Strong server-side state and data-flow discipline - explicit pipeline stage state machine, manifest checkpointing, artifact integrity verification, input fingerprinting, and recovery/invalidations demonstrating deliberate design for correctness and reproducibility.
Evidence
chronica-ai-video-pipeline/src/pipeline/runPipeline.ts: executeStage, persistManifest, fingerprinting, canReuse and invalidation logic
chronica-ai-video-pipeline/src/providers/llm.ts: LlmProvider interface and versioned prompt usage with strict schema and error handling
chronica-ai-video-pipeline/tests/pipeline.test.ts: checkpointing, recovery and artifact-integrity tests demonstrating expected state transitions
UX & Visual Polish
Look and feel quality
Not evidenced in public code
Expertise
Frontend AI Integration• Middle
React• Middle
Industries
Media & Entertainment• Senior
Software• Middle
Technologies
TypeScript• Middle
Node JS• Middle
Tailwind CSS
Next.js
Three.JS
React.js
Vite
hls.js
PostCSS
Zod
React Three Fiber
Recommendations
- Lead development of backend AI media pipelines, LLM integrations and recovery/manifest systems (pipeline orchestration, artifact integrity, checkpointing).
- Implement or own server-side TTS and audio processing features where precise WAV handling and concurrency control are required.
- Build evaluation and monitoring tooling for LLM outputs and pipeline stability (use runEvaluation.ts patterns for latency, stability and adversarial test coverage).
- Collaborate with frontend engineers to connect the proven pipeline backend to production UIs and expose safe, versioned APIs for playback, downloads and manifests.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: