Aizat Murtazin
Python Developer
1+ project
Python
SQL
Go
Model Architecture & Training: 7/10
Experimentation & Evaluation: 7/10
Computational Efficiency: 7/10
Active 20 days ago
+7 (986) 9122280 Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
A competent ML engineer with solid hands-on experience building custom training pipelines and numerical experiment code; the strongest proven skill is implementing and running GPU-batched custom training/evaluation pipelines (e.g., MultiCandidateAngleRNN and oracle optimization in QAOA/angle_rnn_core.py and oracle_lab_core.py). There is limited public evidence of large-scale production deployment, CI/CD for models, or monitoring/drift tooling, and several high-value areas (serving, observability, MLOps automation) are not evidenced in the human-authored code.
Technical skills
Python• Senior • 3y+
SQL• Middle • 3y+
Go• Junior • 1 project
Python
Asyncio
SQLAlchemy
Uvicorn
Pydantic
FastAPI• 3y+
Databases
Apache Kafka
RabbitMQ
Redis
SQLite
Databases
PostgreSQL• 3y+
AI/ML
Airflow
CatBoost
RAG
NumPy
Pandas
Deep Learning
Speech Recognition
Transformers
Jupyter Notebook
Bitsandbytes
LLM• 3y+
DevOps
CI/CD
Git
Grafana
Kubernetes
Prometheus
Rest API
Docker• 3y+
Analytics
Seaborn
QA
Pytest
Selenium
Timeline
Python Developer
•
Middle
Akbars Bank
•
Full-Time
Developed production services for monitoring ML models, including a data collection module used by multiple company models and improvements to model quality metrics. Designed and built ETL pipelines to fetch, validate, and load data from external REST APIs, internal databases, and web sources, automating workflows and reducing processing time. Worked on internal backend modules (profiles, authentication) and improved database performance and API response times; also contributed to an LLM-based news aggregation service by tuning prompts and generation parameters and supporting evaluation via labeled data.
Python
SQL
FastAPI
PostgreSQL
Docker
LLM
Kazan Federal University
Master's Degree •
software engineer
Senior AI/ML Engineer
Confidence: High ML Engineer
A competent ML engineer with solid hands-on experience building custom training pipelines and numerical experiment code; the strongest proven skill is implementing and running GPU-batched custom training/evaluation pipelines (e.g., MultiCandidateAngleRNN and oracle optimization in QAOA/angle_rnn_core.py and oracle_lab_core.py). There is limited public evidence of large-scale production deployment, CI/CD for models, or monitoring/drift tooling, and several high-value areas (serving, observability, MLOps automation) are not evidenced in the human-authored code.
Model Architecture & Training
7/10
Custom model architectures and training loops are present (RNN and multi-candidate models), with deliberate loss choices and training control (early stopping, diversity term).
Evidence
QAOA/angle_rnn_core.py: AngleRNN, MultiCandidateAngleRNN, compute_loss, mc_train_step, train_one_epoch
QAOA/angle_rnn_core.py: run_training, run_training_mc - full custom training loops, checkpointing and eval
Data Pipeline & Feature Engineering
6/10
Feature engineering and dataset pipelines are implemented (raw feature matrix, feature cache, train/val splits, synthetic data generation and incremental labeling).
Evidence
QAOA/angle_rnn_core.py: build_raw_feature_matrix, build_feature_cache, fit_encoder_scaler
QAOA/notebooks/qaoa_teacher_labeling.ipynb: synthetic data generation and teacher labeling loop (records_to_npz, persist_good_record)
Experimentation & Evaluation
7/10
Rigorous experimentation and evaluation tooling: history CSVs, checkpointing, evaluation functions, pilot experiments, summary artifacts and saved predictions for reproducibility.
Evidence
QAOA/angle_rnn_core.py: history CSV writing, checkpoint_{out_prefix}.pt saving, eval_mc_checkpoint and save_best
QAOA/notebooks/qaoa_init_ablation.ipynb: grid runs, checkpointing, matrix/angles CSV and summary generation
MLOps & Deployment
4/10
Basic model lifecycle engineering - checkpointing and artifact saving are present, but there is limited evidence of production serving, CI/CD, monitoring or drift detection for models.
Evidence
QAOA/angle_rnn_core.py: torch.save checkpoints and ART artifact directory usage
QAOA/notebooks: saving CSV/NPZ artifacts and checkpoint files for experiments
Computational Efficiency
7/10
Clear efficiency work - batched GPU evaluation, vectorized restart strategies, careful use of torch.no_grad, chunking for large scans, and Adam+LBFGS orchestration for optimization performance.
Evidence
QAOA/oracle_lab_core.py: optimize_restarts with batched h_big and Adam + LBFGS loops
QAOA/notebooks/qaoa_init_ablation.ipynb: batched p=1 grid evaluation with chunking and GPU batching
Research Depth & Innovation
6/10
Substantive research-depth in a niche scientific ML domain (QAOA/oracle), custom numerical methods and experiment design; this shows applied research rather than off‑the‑shelf usage only.
Evidence
QAOA/oracle_lab_core.py: spectral_features, j_eigen_features, optimized oracle for p_ground and domain-specific metrics
QAOA/angle_rnn_core.py: multi-candidate RNN architecture and candidate scoring tied to domain simulator
Expertise
Audio & Speech Processing• Middle
Scientific AI (Life Sciences, Physics, Climate)• Senior
Industries
Artificial Intelligence• Senior
Science & Engineering• Senior
Technologies
SQL• Middle • 3y+
PostgreSQL• 3y+
Redis
Rest API
CatBoost
Airflow
SQLAlchemy
RabbitMQ
FastAPI• 3y+
Prometheus
Speech Recognition
CI/CD
Pandas
Git
SQLite
Docker• 3y+
Kubernetes
Apache Kafka
Grafana
LLM• 3y+
RAG
Asyncio
Recommendations
- Lead development of research-to-production model training pipelines and experiment reproducibility (training loops, checkpointing, eval artifacts) for scientific ML workloads.
- Implement end-to-end MLOps around existing experiments - model registry, CI for training/eval, reproducible launches and inference-serving stubs.
- Extend audio/ASR work into production-grade services: hardened FastAPI endpoints, containerized deployment, and monitoring for audio ingestion and transcription quality.
- Package and document experiment reproducibility (requirements, run scripts, minimal Docker/Colab recipes) so others can reproduce oracle and RNN experiments reliably.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Data Scientist
Confidence: Medium ML Practitioner
A focused ML researcher/practitioner who builds and evaluates custom RL-for-code training loops and experimental evaluation harnesses. The strongest proven skill is hands-on model training and experimental reward engineering, evidenced by the custom PF-PPO/GRPO training code and reward/evaluation functions in the ASE notebooks (pf-ppo-asem.ipynb, grpo-asem.ipynb). What is not evidenced is productionization - there are no pinned environments, CI-driven reproducible pipelines, or business/impact framing in the public artifacts.
Statistical Rigor
2/10
Limited statistical rigor: custom rewards and aggregated metrics are present, but there are no hypothesis tests, uncertainty quantification, multiple-comparison controls, or documented assumption checks.
Data Wrangling & Cleaning
5/10
Solid basic data wrangling for ML experiments - dataset loading, deterministic split and Dataset wrapper are implemented, plus a pragmatic test harness - but no advanced provenance, DVC, or explicit leakage / label-leakage checks.
Exploratory Analysis & Visualization
2/10
Exploratory analysis is minimal: notebooks print and compute aggregate metrics but lack narrative EDA, diagnostic visualizations, or written interpretations of findings.
Predictive Modeling
6/10
Strong evidence of predictive-model engineering skills for research prototypes: custom PF-PPO/GRPO training loops, pairwise PPO loss, ref-model KL handling, and evaluation hooks are implemented. Missing are production-grade CV pipelines, systematic hyperparameter search, and robust experiment tracking.
Business Insight & Impact
1/10
No evident business-impact framing or cost-of-error analysis in the human-authored artifacts; the work is research/experiment-focused without FP/FN cost discussion or actionable product metrics.
Reproducibility & Notebook Hygiene
3/10
Basic reproducibility steps present (seed, explicit dataset paths, notes in cells) but missing pinned environment / lockfiles, CI integration, data versioning, or modularized experiment pipelines.
Expertise
Analytics• Middle
Technologies
Deep Learning
Python• Senior • 3y+
Bitsandbytes
Jupyter Notebook
Transformers
NumPy
Recommendations
- Prototype and evaluate RL-based model improvements or reward functions for research experiments (PF-PPO/GRPO style training loops).
- Implement LLM evaluation harnesses and local judge/reward integrations for ML research and controlled ablation studies.
- Work on research-to-prototype tasks: integrate model checkpoints, design per-task metrics, and run reproducible experiment sweeps; avoid relying on notebook-only workflows.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Junior Backend Developer
Confidence: High API Engineer
A practical backend-focused developer at a solid early-middle level who builds maintainable REST services and integration glue. The strongest proven skill is building REST API endpoints with authentication and server-side processing for ASR - for example, the upload->transcribe->persist flow implemented across backend/app/main.py and backend/app/asr.py. There is limited evidence of advanced production practices such as migration histories, background workers/queues, systematic observability, or multi-service architecture in the analyzed human-authored code.
API Design
4/10
Clear REST endpoints with consistent HTTP error usage and role checks, but no API versioning, no pagination conventions, and no idempotency strategy beyond single-endpoint logic.
Evidence
pp-voice-commands/backend/app/main.py: multiple /api/* route handlers (login, /api/voice/upload, /api/voice/records, etc.) showing request/response patterns and HTTPException usage
pp-voice-commands/frontend/src/api.ts: authFetch wrapper and client-side usage of token-based auth (shows expected error contract and client contract)
Data Layer & Database
4/10
Uses SQLAlchemy ORM correctly with joinedload to avoid N+1 and clear CRUD boundaries, but no migration history or explicit schema-evolution artifacts visible in the analyzed files.
Evidence
pp-voice-commands/backend/app/crud.py: joinedload(models.VoiceRecord.user), query filters and update patterns
pp-voice-commands/backend/app/main.py: on_startup calls Base.metadata.create_all(bind=engine) - demonstrates DB bootstrapping pattern
Scalability & Performance
2/10
Implementation is synchronous and pragmatic (ASR + ffmpeg), with careful per-request audio conversion and temporary-file cleanup, but no caching, no background/queue-based decoupling, and no rate-limiting or measured performance tuning.
Evidence
pp-voice-commands/backend/app/asr.py: _ensure_wav_16k_mono and _convert_to_wav_16k_mono use ffmpeg subprocess with temp files and cleanup
pp-voice-commands/backend/app/main.py: synchronous endpoints and direct transcribe_file call on upload (no background worker)
System Architecture
3/10
Reasonable module separation (auth, ASR, CRUD, parser, config), showing a small monolith with intentional separation of concerns, but no multi-service decomposition or cross-service contract design.
Evidence
pp-voice-commands/backend/app/asr.py: isolated ASR module encapsulates model resolution and transcription
pp-voice-commands/backend/app/crud.py and pp-voice-commands/backend/app/auth.py: clear separation of persistence and auth logic
Security & Auth
5/10
Good basic security practices - hashed passwords, JWT-based tokens, role checks and path-safety for file access - but some insecure patterns remain (hard-coded example admin seeding, no refresh/revocation strategy, no secrets management evidence).
Evidence
pp-voice-commands/backend/app/auth.py: password hashing and JWT token creation (authentication primitives)
pp-voice-commands/backend/app/main.py: role checks in endpoints and Path(filename).name to avoid path-traversal when serving audio
Reliability & Observability
3/10
Some reliability care visible - logging, try/finally cleanup, DB session close - but no systematic timeouts/retries, no graceful shutdown hooks beyond startup, and limited structured observability/metrics.
Evidence
pp-voice-commands/backend/app/asr.py: logging of ASR warnings/info and try/finally to delete temp work files
pp-voice-commands/backend/app/main.py: on_startup end-to-end DB seeding wrapped with DB close in finally
Verified artifacts
Expertise
Python• Junior
Technologies
Databases
Pydantic
Uvicorn
Recommendations
- Assign to build or extend small-to-medium REST APIs (auth, file uploads, CRUD) where clear boundaries and role checks are required.
- Put on projects integrating local media processing (ASR/ffmpeg) or other CPU-bound external tools that are simpler as synchronous endpoints.
- Have them harden security and operational controls: implement migrations, secrets management, refresh/revocation for tokens, and add structured logging/metrics.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: