Michael Ivkin
AI Engineer
4+ years exp
JavaScript
Python
TypeScript
Model Architecture & Training: 4/10
Active 3 days ago
+7 (911) 9337554
Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
A backend-focused API engineer at a mid-to-senior level who builds production REST services for an education/assessment platform. The strongest proven skill is designing robust exam/variant APIs with transactional submit logic and idempotency controls, demonstrated in src/examgen/api/routers/variants.py (generate/submit paths, transactional checks and ownership guards). Public code lacks explicit large-scale ops artifacts - e.g., comprehensive caching/invalidation strategy, formal load-testing reports, and cluster-level resiliency patterns are not evidenced.
Technical skills
JavaScript
Python• Senior • 4y+
TypeScript• Middle
Python
Pydantic
Asyncio
Celery• 4y+
Django• 4y+
FastAPI
Databases
PostgreSQL• 4y+
Redis• 4y+
ClickHouse
AI/ML
Fine-tuning
Prompt Engineering
PyTorch
Torchvision
NumPy
Pandas
CatBoost
SHAP
Jupyter Notebook
OpenAI SDK
Deep Learning
Computer Vision
Claude
LLM
RAG
Frontend
Next.js
React.js
DevOps
Docker• 4y+
Rest API• 4y+
CI/CD
Docker Compose
Git
GitHub Actions
Kubernetes
Nginx
Analytics
Matplotlib
Seaborn
Timeline
Founder, AI Product Engineer
•
Executive
Ivelika
•
Full-Time
Built an adaptive math practice product end-to-end, including the core engine, product logic, and monetization. Implemented a Bayesian-based adaptive generation approach and addressed cold-start by calibrating priors from exam statistics. Added deterministic generation with reproducibility checks integrated into CI and supporting infrastructure for the product.
Next.js
React.js
TypeScript
FastAPI
PostgreSQL
Docker Compose
Nginx
GitHub Actions
CI/CD
Git
AI Product Engineer
•
Middle
AI Product (Telegram)
•
Full-Time
Developed an AI-powered web app inside Telegram for a large user base, focusing on safety against multi-step prompt-injection attacks. Built a resilient bulk messaging system for millions of users using asynchronous workers and failure recovery. Designed and deployed an on-chain AI agent workflow with task execution and payments escrow, and optimized infrastructure costs using Kubernetes and ClickHouse.
Python
FastAPIsince 2025
Claude
LLM
RAG
PostgreSQL
Redis
ClickHouse
Kubernetes
Docker
Peter the Great Saint Petersburg Polytechnic University
Bachelor's Degree •
Technical physics; Semiconductor photonics and nanoelectronics
Backend Developer
•
Middle
Freedom Developer
•
Full-Time
Worked on payment and integration flows by connecting acquiring and payment channels and processing webhook events. Integrated external systems (e.g., CRM and 1C) and implemented background synchronization to avoid blocking API requests. Built a customer support bot backed by the client knowledge base using LLM capabilities to handle a large share of typical requests.
Pythonsince 2022
Django
PostgreSQLsince 2022
Celery
Redissince 2022
Rest API
Senior AI/ML Engineer
Confidence: High ML Engineer
A senior-level backend and ML prototyping engineer who combines production-grade FastAPI service design with hands-on CNN model implementations in notebooks. The strongest proven skill is building reliable, auditable, transactional backend features for an education product, evidenced by the auth, variant generation, assignment and submission flows (see routers like auth.py and variants.py implementing transactions, idempotency and audit). What is not evidenced is production ML lifecycle work and large-scale MLOps such as model serving, profiling, orchestrated training pipelines or experiment tracking.
Model Architecture & Training
4/10
How well models are designed and trained
Custom CNN architectures and end-to-end training loops are implemented in notebooks (LeNet, AlexNet, VGG, NiN) with explicit forward/initialize_weights and train/test loops on torchvision datasets.
Evidence
digital-image-processing-2024/cnn-intro/lenet.ipynb: class LeNet and train_step/test_step
digital-image-processing-2024/advanced-cnn/vgg-11-16.ipynb: class VGG and training loop
digital-image-processing-2024/advanced-cnn/alexnet.ipynb: class AlexNet and training loop
Data Pipeline & Feature Engineering
3/10
How data is prepared for models
Basic data preprocessing and dataset handling appear (torchvision transforms, DataLoader, KFold and CatBoost preprocessing in a regression notebook) but no large-scale ETL pipelines or production feature stores.
Evidence
digital-image-processing-2024/advanced-cnn/*: use of torchvision transforms and DataLoader
digital-image-processing-2024/house-prices/house-prices-prediction.ipynb: preprocess() with numeric/categorical handling and KFold
Experimentation & Evaluation
3/10
How results are measured and tested
Experimentation includes training/validation loops, cross-validation, plotting and use of SHAP for feature analysis but lacks experiment tracking, reproducible run configs, or systematic ablation harnesses.
Evidence
digital-image-processing-2024/*: train_step/test_step loops and epoch plotting in notebooks
digital-image-processing-2024/house-prices/house-prices-prediction.ipynb: KFold CV and SHAP analysis
MLOps & Deployment
1/10
How models are shipped to production
Production backend shows robust engineering for an edtech service but there is little evidence of model serving, CI model lifecycle, deployment pipelines, or drift monitoring for ML artifacts.
Evidence
ivelika-back/src/examgen/api/services/llm_pipeline.py: presence of lightweight LLM pipeline helpers
ivelika-back/src/examgen/api/routers/health.py: LLM health check endpoints (minimal serving/ops signals)
Computational Efficiency
1/10
How efficiently computing resources are used
Standard batching via DataLoader and reasonable training settings are used, but there is no evidence of GPU profiling, quantization, advanced batching, or memory/throughput optimization.
Evidence
digital-image-processing-2024/*: DataLoader usage and batch_size settings in notebooks
Research Depth & Innovation
1/10
Depth of research and new ideas
Implements classic CNN architectures from scratch but does not reproduce or extend recent research innovations or present novel algorithms or rigorous ablation studies.
Evidence
digital-image-processing-2024/advanced-cnn/*: implementations of canonical architectures (AlexNet, VGG, NiN)
Expertise
Computer Vision & Image Analysis• Middle
Industries
Education• Middle
Technologies
Redis• 4y+
Rest API• 4y+
Claude
CatBoost
ClickHouse
Docker Compose
SHAP
GitHub Actions
Fine-tuning
Prompt Engineering
Computer Vision
OpenAI SDK
CI/CD
Django• 4y+
Git
PyTorch
Docker• 4y+
Kubernetes
Nginx
LLM
RAG
Celery• 4y+
Torchvision
CNN• mentioned only
PyTorch• mentioned only
Recommendations
- Lead backend work for education or assessment products where secure transactional flows, idempotency and auditability are required.
- Prototype and evaluate computer vision models and architectures on toy and benchmark datasets, then collaborate with an MLOps engineer to productionize models.
- Contribute to design and implementation of feature engineering and validation pipelines for supervised learning tasks (structured and image), leveraging existing data preprocessing patterns.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Senior Data Scientist
Confidence: Medium Data Engineer
A senior-level backend and analytics engineer with strong production web-service and data-aggregation skills. The strongest proven skill is building robust, product-aware APIs and analytics pipelines - for example the consolidated student progress builder and transactional submit flow in src/examgen/api/routers/progress.py and src/examgen/api/routers/variants.py. Limited evidence exists for formal statistical methodology, deployment-grade MLOps, pinned reproducible environments or enterprise-scale streaming/Big Data pipelines.
Statistical Rigor
4/10
Correct use of statistics
Evidence of probabilistic student-modeling (BKT) and time-aware trend detection exists, but there are few formal statistical tests, uncertainty quantification, or causal analyses.
Evidence
src/examgen/api/routers/progress.py: uses BKT functions (compute_task_display_level, detect_variant_trend) and builds mastery estimates
src/examgen/api/routers/analytics.py: endpoints for error analysis and task similarity indicating analytic pipelines
Data Wrangling & Cleaning
6/10
Preparing and cleaning data
Strong production-grade data handling and defensive logic - event aggregation, transactional answer submission, draft handling, timezone and idempotency safeguards are implemented.
Evidence
src/examgen/api/routers/progress.py: build_student_progress aggregates event storage, computes rolling/overall accuracy and belief maps
src/examgen/api/routers/variants.py: submit_variant implements transactional batch checking, assignment locking and idempotency guards
src/examgen/api/routers/assignments.py: _enrich_assignment uses variant storage and event counts with defensive logging
Exploratory Analysis & Visualization
3/10
Exploring and visualizing data
Notebooks include many visualizations and plots but are exploratory, toy-dataset oriented and lack written interpretation or formal data-story scaffolding.
Evidence
digital-image-processing-2024/advanced-cnn/vgg-11-16.ipynb: training plots and activation visualization
digital-image-processing-2024/cnn-intro/lenet.ipynb: activation map visualization and training traces
Predictive Modeling
3/10
Building models that predict
There are hands-on predictive experiments (CNNs, CatBoost) and baseline training loops, but models are research/toy experiments rather than productionized predictive systems with rigorous CV, calibration, or deployment artifacts.
Evidence
digital-image-processing-2024/house-prices/house-prices-prediction.ipynb: CatBoost KFold training and SHAP analysis
digital-image-processing-2024/advanced-cnn/alexnet.ipynb: custom CNN training loops on FashionMNIST
Business Insight & Impact
4/10
Turning analysis into business value
Product-aware backend design shows business thinking - tutor/student flows, anti-cheat, rate limiting and idempotent share links - but explicit business-metric analysis or cost-of-error reasoning is limited.
Evidence
src/examgen/api/routers/variants.py: anonymous vs authenticated anti-scraping and submission_mode handling
src/examgen/api/routers/assignments.py: idempotent share link creation and tutor-student authorization checks
src/examgen/api/routers/progress.py: single consolidated progress API shaped for frontend needs
Reproducibility & Notebook Hygiene
2/10
Clean, repeatable analysis
Notebooks exist and some have meta, but environment pinning, data versioning, CI tests or reproducible pipeline artifacts are missing or sparse.
Evidence
digital-image-processing-2024/advanced-cnn/vgg-11-16.ipynb: notebook with '# NOTEBOOK META' but no pinned environment or requirements.txt
digital-image-processing-2024/house-prices/house-prices-prediction.ipynb: dead_notebook/never_executed flag in metadata
Expertise
Analytics• Middle
Industries
Education• Middle
Technologies
Deep Learning
Jupyter Notebook
Seaborn
Matplotlib
Pandas
NumPy
CNN• mentioned only
PyTorch• mentioned only
Recommendations
- Lead backend work for education analytics services - design and implement event-aggregation, BKT updates and tutor-facing APIs (use existing progress.py and variants.py patterns).
- Implement reproducible ML pipelines and MLOps - add environment pins, CI runs for notebooks converted to formal training scripts and data versioning for the house-prices and CNN experiments.
- Productize predictive components carefully - convert experimental notebooks into tested libraries with proper CV, calibration and monitoring before deployment.
- Expand observability and testing for analytics - add unit/integration tests around BKT updates, add load tests for variant generation and transactional submit paths.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Senior Backend Developer
Confidence: Medium API Engineer
A backend-focused API engineer at a mid-to-senior level who builds production REST services for an education/assessment platform. The strongest proven skill is designing robust exam/variant APIs with transactional submit logic and idempotency controls, demonstrated in src/examgen/api/routers/variants.py (generate/submit paths, transactional checks and ownership guards). Public code lacks explicit large-scale ops artifacts - e.g., comprehensive caching/invalidation strategy, formal load-testing reports, and cluster-level resiliency patterns are not evidenced.
API Design
7/10
How well APIs are designed
API surfaces are well-structured with pydantic models, consistent error handling, rate-limiting, and explicit idempotency/authorization checks; versioning middleware and response-stripping for anti-scraping show deliberate API design.
Evidence
ivelika-back/src/examgen/api/routers/variants.py: generate_variant / submit_variant endpoints with Pydantic request/response models and detailed HTTPException usage
ivelika-back/src/examgen/api/app.py: APIVersionMiddleware dispatch (API versioning middleware present)
ivelika-back/src/examgen/api/routers/variants.py: _return_variant_response uses answer-stripping for anonymous/deferred responses (anti-scraping / id exposure control)
Data Layer & Database
7/10
Working with databases
Clear transactional control, connection-pool usage and explicit row locking for race guards, combined with storage classes and careful snapshot persistence indicate deliberate DB design and transaction reasoning.
Evidence
ivelika-back/src/examgen/api/routers/variants.py: submit_variant uses storage.connection() with conn.transaction(), for_update locks on assignment rows and sequential BKT updates
ivelika-back/src/examgen/assignments/storage.py: AssignmentStorage with AsyncConnection/psycopg_pool usage and explicit methods like mark_started, update_assignment_status
pyproject.toml: dependencies include psycopg[binary] and psycopg-pool (evidence of Postgres + connection pooling usage)
Scalability & Performance
5/10
Handling load and speed
Some scalability and performance patterns are present (rate limiting, connection pooling, non-blocking snapshot saves), but no large-scale distributed tuning, explicit caching-invalidation strategy, or documented load testing artifacts were found.
Evidence
ivelika-back/src/examgen/api/rate_limit.py: rate limiter helpers and trusted networks parsing
ivelika-back/src/examgen/api/routers/variants.py: use of Async snapshot saves guarded by try/except to avoid blocking generation path
pyproject.toml and use of psycopg-pool in several modules (e.g., assignments/storage.py) for pooled DB connections
System Architecture
6/10
Overall system structure
Modular decomposition (api routers, services, assets, variants, snapshots) and plugin-like registry for generators show deliberate architecture choices; boundaries between API, generation, and storage are clear.
Evidence
ivelika-back/src/examgen/api/routers/__init__.py: many well-separated routers (variants, practice, assignments, etc.) showing service decomposition
ivelika-back/src/examgen/api/services/generation.py and api/services/llm_pipeline.py: separation of generation and LLM-specific pipelines
ivelika-back/src/examgen/assets/* and examgen/core/registry: asset/layout/render subsystems and registry pattern for generators
Security & Auth
6/10
Protecting data and access
Strong auth and authorization patterns (token decoding, require_tutor, ownership checks), input validation via Pydantic, and rate limiting; some secret/rotation lifecycle details are not visible in code snippets but cookie attributes and masking utilities are present.
Evidence
ivelika-back/src/examgen/auth/dependencies.py: token extraction, get_current_user, require_tutor and ownership verification hooks
ivelika-back/src/examgen/auth/cookies.py: cookie attribute helpers and secure-flag logic
ivelika-back/src/examgen/api/routers/variants.py: owner/assignment checks (_can_view_variant) and rate-limiter usage
Reliability & Observability
5/10
Stability and monitoring
Reliability patterns (idempotency handling, transaction-level race guards, defensive logging) are present, plus structured logging calls; fewer visible patterns for retries with jitter, circuit breakers, or a comprehensive observability stack in the analyzed files.
Evidence
ivelika-back/src/examgen/api/routers/variants.py: idempotency path for already-submitted variants and transactional save with deletion of draft in the same transaction
ivelika-back/src/examgen/api/routers/variants.py: logger.exception / logger.warning usage around generation and snapshot saves
roko_acp/logger.py: JsonFormatter and request id / contextvar helpers (structured logging primitives present)
Expertise
Python• Middle
Backend AI & LLM• Middle
Microservices & API Architecture• Middle
Technologies
Python• Senior • 4y+
PostgreSQL• 4y+
FastAPI
Asyncio
Pydantic
Recommendations
- Use to implement and evolve REST microservices that require careful DB transactions, idempotent operations, and authorization checks (e.g., answer submission, assignment lifecycle).
- Build and extend exam/variant generation pipelines and LLM-integration points (the developer has LLM pipeline and generator patterns already wired).
- Improve reliability and observability by adding structured metrics, retry-with-backoff for external calls, and documented caching/invalidation (they can own the end-to-end upgrade).
- Work on feature development that requires strong API contracts between frontend and backend (deferred drafts, draft restore, and anti-cheat answer-stripping flows).
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: