Igor Afonin
AI Engineer
4+ years ML exp
Management: Under a year
Python
TypeScript
Node JS
SQL
JavaScript
Experimentation & Evaluation: 7/10
Computational Efficiency: 7/10
MLOps & Deployment: 6/10
Invite to interview
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
A senior-level LLM-engineering practitioner focused on trustworthy LLM evaluation, agent workflows, and inference-capacity tooling. The strongest proven skill is building rigorous evaluation and benchmarking pipelines rooted in deterministic validity gates and honest scoring (e.g., scorecard construction in assay/assay/scorecard.py and solver orchestration in assay/assay/solver.py). What is not evidenced is custom model training or novel neural-architecture development—there are analytical and simulation layers but no bespoke training loops or new model implementations in public code.
Phone
Technical skills
Languages
5
Python
TypeScript
Node JS
SQL
JavaScript
Python
5
FastAPI
Pydantic
Asyncio
Uvicorn
aiogram
AI/ML
42
LLM
Claude
NumPy
vLLM
SciPy
Computer Vision
NLP
AutoGen
CrewAI
DeepSeek
DSPy
Gemini
LangGraph
Llama
LlamaIndex
Mistral
PyTorch
Ray
Scikit-learn
XGBoost
LangChain
Transformers
LoRA
AWQ
CatBoost
Composio
ControlNet
DeepEval
GPTQ
LightGBM
PEFT
Pydantic AI
QLoRA
Qwen
Stable Diffusion
BERT
CUDA Toolkit
GigaChat
MLFlow
SGLang
Deep Learning
MLOps
Frontend
6
React.js
Next.js
Tailwind CSS
React Query
Zod
Bootstrap
Databases
8
PostgreSQL
SQLite
MySQL
Redis
ClickHouse
FAISS
pgvector
Qdrant
DevOps
7
AWS
GCP
Kubernetes
Amazon EC2
Cloudflare
OpenTelemetry
Google GKE
Other
19
Jest
Pytest
Power BI
Langfuse
LangSmith
Perplexity
RunPod
Docker
Rest API
Postman
PoC Library
CI/CD
RAG
Fine-tuning
KEDA
Vector
SLI/SLO/SLA
Knowledge Distillation
NER
Timeline
Founding AI Engineer / AI Platform Architect
•
Lead
AI Platform (NDA)
•
Cofounder
Built a no-code AI agent platform for multi-tenant production use with continuous availability. The CV indicates responsibility for core platform design and operations supporting multiple customers and many runtime endpoints. Specific tools and frameworks are not detailed for this role.
Lead AI Engineer
•
Lead
NeuroChief
•
Full-Time
Led AI engineering activities for enterprise deployments and managed a small team of engineers. The CV emphasizes delivery of multiple enterprise rollouts within the stated time window, with responsibilities centered on technical leadership and execution. Specific technologies used are not listed for this role.
AI Engineer
•
Middle
AI MakeBiz
•
Full-Time
Worked as an AI engineer on ML/AI initiatives during this period. The CV does not provide specific responsibilities, technologies, or outcomes for this job entry. It is listed as part of the candidate’s progression into LLM/agent-focused work later in career.
ML/AI Engineer
•
Middle
Devias (SprintLabs)
•
Full-Time
Started the career in classic ML and data science and progressed toward deep learning, early LLM, and agent-oriented prototypes. The CV states multiple PoC to MVP transitions across products during this timeframe. Specific tools and libraries for this role are not enumerated in the job entry.
Senior AI/ML Engineer
Confidence: High LLM Engineer
A senior-level LLM-engineering practitioner focused on trustworthy LLM evaluation, agent workflows, and inference-capacity tooling. The strongest proven skill is building rigorous evaluation and benchmarking pipelines rooted in deterministic validity gates and honest scoring (e.g., scorecard construction in assay/assay/scorecard.py and solver orchestration in assay/assay/solver.py). What is not evidenced is custom model training or novel neural-architecture development—there are analytical and simulation layers but no bespoke training loops or new model implementations in public code.
Model Architecture & Training
1/10
How well models are designed and trained
Minimal-to-no evidence of custom model architectures or training loops; the code uses hosted models and model-family stamping/judging rather than building or training models from scratch.
Evidence
assay/assay/draft_agent.py: SdkAgentAnalyzer.directionality and model_family usage (generator/judge/sut stamping)
assay/assay/score/t3.py: judge orchestration and judge prompt build (judge as external model consumer rather than custom architecture)
Data Pipeline & Feature Engineering
4/10
How data is prepared for models
Clear, tested data-handling for evaluation corpora, fixture parsing and derivation; workload priors and corpus construction are explicit and guarded by validation.
Evidence
assay/assay/dataset.py: parse_fixture, load_fixtures, fixture_to_sample — fixture parsing and validation
llmops-cc/llmops_calc/llmops_calc/concurrency/workload_loader.py: load_workload_priors — workload priors/data loading
Experimentation & Evaluation
7/10
How results are measured and tested
Strong experimentation and evaluation engineering - comprehensive harness, scorer/scorecard, statistical pass^k computations, benchmark logging and analysis pipelines with unit tests and deterministic gating.
Evidence
assay/assay/scorecard.py: build_scorecard, _epoch_vectors, render and write_results_md — scorecard construction and honesty rules
edp/benchmark/logger.py: manifest logging, write/open manifest, and structured run metadata
assay/tests/unit/* and llmops-cc/tests/unit/*: extensive unit tests exercising contract properties and statistical assertions
MLOps & Deployment
6/10
How models are shipped to production
Substantial MLOps and deployment engineering: CLI tooling, precondition checks, RunPod integrations, MCP adapters, and guarded live-run semantics (budget, watchdogs, freeze/pin), plus MCP/server adapters.
Evidence
llmops-cc/lib/runpod_client.py: RunPod pricing and API interactions
llmops-cc/lib/preconditions/_common.py: precondition checks for environment, GPU, network, disk and other runtime prechecks
assay/assay/solver.py: atomic reserve budgeting, watchdog, and integration with inspect_ai.solver for orchestration
Computational Efficiency
7/10
How efficiently computing resources are used
Significant work on computational efficiency and inference modeling - roofline/attention models, KV-cache calculations, batching and Monte Carlo capacity simulation, with reproducible analytical outputs and profiling hooks.
Evidence
llmops-cc/llmops_calc/llmops_calc/analytical/roofline.py: roofline classification and time_per_layer formulas
llmops-cc/llmops_calc/llmops_calc/concurrency/monte_carlo.py: end-to-end Monte Carlo capacity simulation and batching composition
llmops-cc/lib/profile/gpu.py and lib/profile/kernel.py: GPU and kernel probing/parsing utilities
Research Depth & Innovation
6/10
Depth of research and new ideas
Research-informed engineering: paper-cited formulas, arXiv citations surfaced in outputs, and implementations of attention variants and KV/eviction heuristics; not a novel research paper but careful reproduction and integration of SOTA formulas.
Evidence
llmops-cc/llmops_calc/llmops_calc/analytical/attention/*: MHA/GQA/MLA/SWA/hybrid attention formulas and dispatch
llmops-cc/llmops_calc/llmops_calc/whatif/* and tests: what-if sweeps and Pareto analysis referencing academic sources
Expertise
AI Agents & Agentic Workflows• Senior
AI Infrastructure & Optimization• Senior
MLOps & Model Lifecycle• Senior
LLM• Senior
RAG• Middle
Industries
Artificial Intelligence• Senior
Data & Analytics• Senior
Software• Senior
Technologies
Deep Learning
MLOps
Python• Senior
SQL
MySQL
Redis
LangGraph
AutoGen
Rest API
LangChain
Claude
Qwen
CatBoost
GCP
ClickHouse
pgvector
FAISS
Qdrant
DeepSeek
Stable Diffusion
DSPy
LlamaIndex
LoRA
vLLM
CUDA Toolkit
MLFlow
XGBoost
OpenTelemetry
KEDA
Fine-tuning
Scikit-learn
SciPy
Knowledge Distillation
Computer Vision• since 2022
NLP• since 2023
NER
SGLang
AWQ
Composio
ControlNet
DeepEval
GPTQ
LightGBM
Langfuse
LangSmith
PEFT
Pydantic AI
QLoRA
Llama
Mistral
GigaChat
CI/CD
Transformers
NumPy
PyTorch
AWS
Docker
Kubernetes
CrewAI
Cloudflare
Gemini
LLM• since 2023
RAG
Perplexity
BERT
Asyncio
Pydantic
Ray
aiogram
Google GKE
Amazon EC2
Vector
SLI/SLO/SLA
RunPod
AI Agents• mentioned only
Claude Code• mentioned only
RunPod• mentioned only
Recommendations
- Lead development of LLM evaluation harnesses and adversarial-resistance test suites (build on assay: scorecard, antigaming, and solver modules).
- Design and ship inference-capacity and cost-planning tools (extend llmops_calc Monte Carlo and roofline components into service-level recommendations).
- Integrate guarded production agent adapters and decision auditing (EDP adapters and MCP hooks) for enterprise agent deployments.
- Implement monitoring and observability for model inference (profilers, e2e traces, and precondition checks) and automate preflight verifications for deployments.
Senior Backend Developer
Confidence: High API Engineer
A backend-focused API engineer at a Senior level who repeatedly demonstrates sound production-grade BFF and service design. The strongest proven skill is building resilient HTTP/streaming backends with careful auth, idempotency, transactional DB handling and tests (e.g. generate.ts, anon-stream.ts, migrations tests). The public code shows limited evidence of large-scale performance benchmarking or platform-level orchestration (k8s/infra), and some parts of the monorepo are generated/templated rather than original domain logic.
API Design
7/10
How well APIs are designed
API design shows deliberate decisions: idempotency key handling, SSE proxying, clear error/status shapes, and auth/403-vs-404 reasoning with tests.
Evidence
mcpgen/apps/api/src/routes/v1/generate.ts: Idempotency-Key handling, 202 response + sse_url, cache-hit vs fresh path logic
mcpgen/apps/api/src/routes/v1/jobs/anon-stream.ts: authorizeJobAccess + resolveAuthSignals and SSE replay vs engine proxy
mcpgen/apps/api/src/routes/v1/playground.ts: authorized playground endpoints, validators and SSE invoke/persist pattern
Data Layer & Database
7/10
Working with databases
Data layer demonstrates schema migrations with idempotency and failure/restore, explicit transactional claim logic, FK-awareness and tested migration evolution.
Evidence
anchr/tests/migrations.test.ts: exhaustive migration runner tests (idempotency, .bak restore, NOT NULL trap)
anchr/src/storage/migrations.ts: multiple migrations (001..003) and runMigrations with backup/restore
mcpgen/apps/api/src/routes/v1/generate.ts & apps/api/tests/claim-generation.test.ts: DB write patterns and transactional claim behaviour
Scalability & Performance
6/10
Handling load and speed
Scalability and performance patterns present - cache-hit replay, SSE stream draining, rate-limit integration and detached engine drains - but no large-scale load-test artifacts.
Evidence
mcpgen/apps/api/src/routes/v1/generate.ts: cache-hit lookup, createCacheHitChild and detached engine SSE drain (exec.waitUntil)
mcpgen/apps/api/src/routes/v1/jobs/anon-stream.ts: cache-replay timeline and fast <1s replay path
mcpgen/package.json: scripts for observability (outbox:monitor) and workspace-level tooling indicating operational concerns
System Architecture
6/10
Overall system structure
Thoughtful service/module boundaries (BFF, playground, engine, cli), feature-flag integration, and clear separation between anon and authed paths; monorepo structure supports modular decomposition.
Evidence
mcpgen/package.json: workspace layout and cross-package boundaries (apps, packages, infra)
mcpgen/apps/api/src/routes/v1/generate.ts and playground.ts: BFF decomposition and per-path design (anon vs authed)
mcpgen/packages/runtime-sdk & packages/contracts: explicit contract packages and shared-config
Security & Auth
7/10
Protecting data and access
Strong auth/authz attention - cookie-scoped anon sessions, JWT org checks, CSRF tests, careful 403/404 decisions and defensive fallbacks.
Evidence
mcpgen/apps/api/src/routes/v1/jobs/anon-stream.ts: authorizeJobAccess and resolveAuthSignals (cookie vs JWT ownership logic)
mcpgen/apps/api/tests/claim-generation.test.ts: CSRF and claim flow tests validating auth boundary and idempotency
mcpgen/apps/api/src/routes/v1/playground.ts: authorizePlaygroundAccess with ULID/UUID format dispatch and explicit 404-vs-403 policy
Reliability & Observability
7/10
Stability and monitoring
Reliability and observability practices visible - defensive try/catch, fallbacks for DB/engine unavailability, backups/restore for migrations, and best-effort persistence with logging.
Evidence
anchr/tests/migrations.test.ts: tests for migration failure restore (.bak) and MigrationFailedError carrying metadata
mcpgen/apps/api/src/routes/v1/generate.ts: try/catch around engine kickoff, drainEngineStream with warnings and exec.waitUntil integration
mcpgen/apps/api/src/routes/v1/playground.ts: persistRun wraps DB writes in try/catch and logs rather than failing the SSE
Verified artifacts
Expertise
Node.js• Senior
Python• Senior
Microservices & API Architecture• Senior
Backend AI & LLM• Senior
Industries
Artificial Intelligence• Senior
Software• Senior
Technologies
PostgreSQL
FastAPI
SQLite
Uvicorn
Recommendations
- Lead development of HTTP BFFs and API gateways that require idempotency, SSE/stream proxies, and careful auth boundaries.
- Implement DB-backed features needing schema migrations, transactional claim flows, and robust rollback/restore strategies.
- Build LLM orchestration components and evaluation harnesses (engine kickoffs, SSE drains, cache-replay timelines, judge pipelines).
- Design and own medium-scale microservices integration points (feature flags, rate limits, partial-state caching) where API contracts are critical.
Senior Frontend Developer
Confidence: Medium Fullstack
A practical senior full-stack engineer who authors robust backend BFFs, CLI/TUI tooling and careful async flows. The strongest proven skill is designing resilient async/stateful server systems and streaming flows, exemplified by the generate.ts SSE drain + cache-hit child logic and the jobs anon-stream handlers. Public code shows strong backend and CLI craftsmanship but limited evidence of authored web component libraries, comprehensive web accessibility work, or a dedicated frontend component design system.
UI Component Architecture
2/10
How interface parts are built
Limited evidence of UI component architecture: there is a well-factored TUI renderer with composable pure functions but no authored web component library, design system, or React component architecture in the human-authored files.
Evidence
anchr/src/render/card.ts
anchr/src/render/pickerLine.ts
Responsive & Cross-browser
1/10
Works on all screens and browsers
Minimal evidence for responsive or cross-browser frontend work; human-authored files show CLI/TUI and backend code but no responsive layout, RTL, or CSS architecture.
Performance Optimization
5/10
Speed of the interface
Concrete performance engineering and measured decisions are present: SSE streaming and background drains, defensive engine drain with waitUntil, runtime detection for safe SQLite usage, and explicit FTS5 performance planning.
Evidence
mcpgen/apps/api/src/routes/v1/generate.ts
mcpgen/apps/api/src/routes/v1/jobs/anon-stream.ts
anchr/.planning/phases/01-storage-foundation/01-07-PLAN.md
Accessibility & Semantics
1/10
Usable for everyone
Little evidence of web accessibility work or ARIA on custom widgets; the code includes TUI prompt/raw-mode handling but no explicit focus management or keyboard/ARIA patterns for custom web widgets.
Evidence
anchr/src/cli/_prompts.ts
anchr/src/tui/hooks/useTasks.ts
State Management & Data Flow
5/10
Managing data in the app
Strong server- and async-state discipline: idempotent DB updates, careful transactional claim flows, cache-hit child creation, SSE partial-state aggregation and explicit race-condition mitigations demonstrate solid data flow design.
Evidence
mcpgen/apps/api/src/routes/v1/generate.ts
mcpgen/apps/api/src/routes/v1/jobs/anon-stream.ts
mcpgen/apps/api/src/lib/job-partial-state.ts
UX & Visual Polish
4/10
Look and feel quality
Good UX polish for CLI/TUI: careful truncation, ANSI styling, detailed card layout and edge-state rendering; less evidence for polished web UI surfaces in human-authored files.
Evidence
anchr/src/render/card.ts
lucid-context/src/cli.ts
Expertise
React• Middle
Modern Web Frameworks• Middle
Web Performance & Optimization• Senior
Industries
Financial Services• Middle
Software• Middle
Technologies
TypeScript• Senior
Node JS• Senior
Tailwind CSS
Next.js
Bootstrap
React.js
React Query
Zod
Recommendations
- Ownership of BFF/API endpoints that need robust async state (SSE, idempotency, cache-hit replay, transactional claim flows).
- Developing developer tools and CLI/TUI features where careful UX, truncation, ANSI styling and deterministic output matter.
- Implementing backend-for-frontend integrations and playground UIs that connect streaming engine pipelines to web/CLI clients (SSE proxying, partial-result aggregation).
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: