Ruslan Musin

AI/ML • Senior • 31 year • Moscow, Russia • UTC+03:00 • In office • Work permits: Russia • Available now
AI Engineer
SQL
Python
Data Pipeline & Feature Engineering: 6/10
MLOps & Deployment: 5/10
Computational Efficiency: 5/10
Active 18 days ago
+7 (925) 7020886
Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles

Overview

A production-focused LLM/AI engineer at a senior level who designs orchestration and multimodal generation pipelines rather than training new models. The strongest proven skill is building robust generative pipelines and media-oriented orchestration, demonstrated by the SonGo pipeline's subtitle alignment, lip-sync shot planning and multi-provider generation/fallback logic (songo_pipeline/live_services.py and pipeline.py). There is little or no evidence of custom model training, formal experiment tracking, or reproducible training workflows in public code.

Technical skills

SQL
Python• Senior
Python
aiogram
Asyncio
Uvicorn
HTTPX
FastAPI
Databases
PostgreSQL
Supabase
AI/ML
AI Agents
Claude
Claude Code
Cursor
Embeddings
Fine-tuning
Gemini
OpenRouter
Qwen
PEFT
OpenAI SDK
Google GenAI SDK
Whisper
ElevenLabs
LLM
LoRA
PyTorch
RAG
DevOps
Docker
Git
Rest API
Vector

Timeline

AI Engineer (R&D) • Middle
Lofty • Full-Time
Dec 2025 to May 2026 5 Months Moscow In office
Researched and built LLM-based AI solutions and developed AI services/agents for internal company products. Designed backend services in Python with FastAPI and integrated LLMs via APIs to support user workflows. Created an MVP and research prototypes, including an AI support agent for SoundSeed with automated request analysis, response generation, and escalation, reducing typical handling time by 60% and lowering operator workload. Also evaluated an open-source music generation approach using LoRA fine-tuning and GPU training to reach vocal quality comparable to commercial systems.
Python
FastAPI
LLM
LoRA
PyTorch
AI Engineer • Middle
Personal AI Products • Full-Time
Jan 2025 to Dec 2025 11 Months In office
Developed end-to-end AI products based on modern LLM and generative AI, from ideation and architecture through MVP creation, model integration, and launch of working user scenarios. Built an AI avatar/content pipeline that automates ad video production (scripts, voice, video, subtitles, and final assembly) with multiple generative models. Created an emotional fitness assistant using OpenAI API and a RAG-based knowledge base, and implemented an answering pipeline to improve response relevance and reduce hallucinations. Developed a nutrition assistant that analyzes photos and user text via OpenAI API to recognize dishes, store history, and generate personalized recommendations.
LLMsince 2025
RAG
Senior AI/ML Engineer Confidence: High LLM Engineer
A production-focused LLM/AI engineer at a senior level who designs orchestration and multimodal generation pipelines rather than training new models. The strongest proven skill is building robust generative pipelines and media-oriented orchestration, demonstrated by the SonGo pipeline's subtitle alignment, lip-sync shot planning and multi-provider generation/fallback logic (songo_pipeline/live_services.py and pipeline.py). There is little or no evidence of custom model training, formal experiment tracking, or reproducible training workflows in public code.
Model Architecture & Training
2/10
How well models are designed and trained
No custom model training or architecture implementation; code shows engineered use of hosted LLM/Multimodal APIs with prompt engineering and parsing logic but not model-building or training.
Evidence
lighter/services/claude.py: parse_food_text, analyze_photo, get_nutrition - orchestration of OpenAI chat/completions/responses
songo_pipeline/live_services.py: generate_music_track, generate_video_clip - consumption of Gemini/Google GenAI APIs
Data Pipeline & Feature Engineering
6/10
How data is prepared for models
Robust data and media pipelines with careful preprocessing, alignment, and DB persistence; explicit handling of TTLs, idempotency, and audio/video subtitle alignment indicates strong production data-pipeline engineering.
Evidence
lighter/db/database.py: create_meal_draft/pop_meal_draft, begin_processed_update/complete_processed_update (idempotency and TTL handling)
songo-ai-avatar-pipeline/songo_pipeline/pipeline.py: build_subtitle_artifacts, transcribe_with_whisper, align_subtitles_with_whisperx (audio transcription and alignment pipeline)
Experimentation & Evaluation
2/10
How results are measured and tested
Minimal evidence of formal experimentation or evaluation tooling - some runtime metrics and estimated cost bookkeeping exist but no experiment tracking, reproducible training runs, or AB/eval harnesses.
Evidence
songo-ai-avatar-pipeline/songo_pipeline/pipeline.py: manifest estimated_cost_usd and run manifest generation
lighter/api/index.py: admin_metrics and admin_get_usage_metrics exposing basic operational metrics
MLOps & Deployment
5/10
How models are shipped to production
Clear production-grade deployment and orchestration: FastAPI endpoints, startup DB init, webhook handling, scheduler/cron endpoints and provider readiness checks; good operational patterns but no full MLOps stack (no model versioning or drift monitoring).
Evidence
lighter/api/index.py: FastAPI app with webhook, cron endpoints, startup init_db and admin dashboard
songo-ai-avatar-pipeline/songo_pipeline/live_services.py: provider_ready and provider-based fallbacks, save_creative_outputs for artifact provenance
Computational Efficiency
5/10
How efficiently computing resources are used
Practical attention to efficiency and resource selection: async DB pool sizing, device selection for Whisper/whisperx, caching of alignment models and ffmpeg usage; focused on pragmatic performance rather than low-level GPU/kernel optimizations.
Evidence
lighter/db/database.py: get_pool uses asyncpg.create_pool with min_size/max_size and statement_cache_size=0
songo-ai-avatar-pipeline/songo_pipeline/pipeline.py: subtitle_ml_device and WHISPERX_ALIGN_CACHE for model/device selection and caching
Research Depth & Innovation
5/10
Depth of research and new ideas
Strong applied research / productization of multimodal generation flows and original heuristics (subtitle allocation, lip-sync phrase extraction, shot planning); not academic-level novel algorithms but substantial original engineering.
Evidence
songo-ai-avatar-pipeline/songo_pipeline/live_services.py: build_fallback_shots, normalize_shot_plan, build_lip_sync_shots_from_lyrics (original multimodal generation heuristics)
songo-ai-avatar-pipeline/songo_pipeline/pipeline.py: allocate_subtitles_across_ranges, subtitle_highlight_mode, distribute_words_in_range (subtitle timing and weighting algorithms)
Expertise
LLM• Senior
Industries
Media & Entertainment• Senior
Software• Middle
Lifestyle• Middle
Technologies
SQL
Supabase
Cursor
Rest API
Claude
Qwen
Claude Code
LoRA
ElevenLabs
Fine-tuning
Embeddings
AI Agents
Google GenAI SDK
OpenAI SDK
PEFT
Git
PyTorch
Docker
Gemini
LLM
RAG
Whisper
Vector
OpenRouter
Flask• mentioned only
Recommendations
  • Lead development of LLM-driven content-generation services (multimodal pipelines that integrate text, audio, and video providers).
  • Implement production integrations where deterministic prompt->artifact flows, provider-fallbacks and artifact provenance are required.
  • Build developer-facing tooling for subtitle/timing and lip-sync workflows or embed the existing heuristics into an SDK.
  • Extend MLOps capabilities by adding experiment tracking, model/version metadata and drift/quality monitoring for generated outputs.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
lighter songo-ai-avatar-pipeline
Senior Backend Developer Confidence: Medium API Engineer
A senior backend engineer focused on Python services and AI integrations with production-grade bot and media-pipeline experience. The strongest proven skill is building AI orchestration and media pipelines, evidenced by songo_pipeline/pipeline.py and songo_pipeline/live_services.py which implement provider abstraction, fallbacks, and manifested run outputs. Public code shows limited formal testing, CI/CD artifacts, and limited evidence of systematic retry/backoff or distributed tracing.
API Design
5/10
How well APIs are designed
Practical, production-ready API surface with secured webhook endpoints, admin UI, and clear secret checks; lacks formal API versioning policy and contract tests but demonstrates deliberate endpoint design and webhook/tokenized routes.
Evidence
ruslanmusin-ai/lighter/api/index.py:telegram_webhook (validates TELEGRAM_WEBHOOK_SECRET and feeds aiogram Update) ruslanmusin-ai/lighter/api/index.py:cron_digest and cron_subscription_renewals (Bearer CRON_SECRET auth checks) ruslanmusin-ai/lighter/api/index.py:setup_webhook and yookassa_webhook_tokenized (internal secret gating and tokenized webhook route)
Data Layer & Database
6/10
Working with databases
Robust async data layer with explicit connection pool configuration, extensive parameterized queries, upserts and idempotence mechanics, and a lease-based processed_updates pattern; missing visible multi-step transaction demarcation or explicit isolation-level tuning.
Evidence
ruslanmusin-ai/lighter/db/database.py:get_pool (asyncpg.create_pool with min_size/max_size and statement_cache_size) ruslanmusin-ai/lighter/db/database.py:save_billing_payment (INSERT ... ON CONFLICT upsert with idempotence_key handling) ruslanmusin-ai/lighter/db/database.py:begin_processed_update / complete_processed_update (lease-based processed_updates to avoid double-processing)
Scalability & Performance
4/10
Handling load and speed
Async code and connection pooling are used and the pipeline includes provider fallbacks and background schedulers; there is little evidence of caching strategies, rate limiting, formal load-testing, or systematic backoff/retry policies for external calls.
Evidence
ruslanmusin-ai/lighter/db/database.py:get_pool (connection pool tuning) ruslanmusin-ai/lighter/bot.py (uses apscheduler AsyncIOScheduler for background jobs) ruslanmusin-ai/songo-ai-avatar-pipeline/songo_pipeline/pipeline.py:generate_audio_track and generate_video_clip (provider orchestration with fallback paths)
System Architecture
6/10
Overall system structure
Clear modular architecture and stage-based pipeline orchestration with provider abstraction, fallback behaviors, progress events and manifest outputs; shows deliberate separation between bot handlers, services, and data layer supporting graceful fallback and operational visibility.
Evidence
ruslanmusin-ai/songo-ai-avatar-pipeline/songo_pipeline/pipeline.py:run_pipeline / resolve_providers (stage orchestration and provider abstraction) ruslanmusin-ai/songo-ai-avatar-pipeline/songo_pipeline/live_services.py:generate_creative_bundle and merge_creative_bundle (provider + fallback merging and artifact saving) ruslanmusin-ai/lighter/api/index.py:on_startup and admin_dashboard (separation of admin API, cron endpoints, and startup DB init)
Security & Auth
5/10
Protecting data and access
Good security hygiene at API boundaries including webhook token checks, internal secret gating and HTML escaping; some security concerns remain such as no public dependency/audit evidence and no full auth lifecycle (refresh/revocation) visible in the code.
Evidence
ruslanmusin-ai/lighter/api/index.py:_require_internal_secret and _require_cron_secret (internal API secret and cron bearer validation) ruslanmusin-ai/lighter/api/index.py:telegram_webhook (validates X-Telegram secret header) and yookassa_webhook_tokenized (token route) ruslanmusin-ai/lighter/db/database.py (parameterized queries using asyncpg to avoid SQL injection)
Reliability & Observability
4/10
Stability and monitoring
Basic reliability and observability practices present - structured logging, try/except with exception logs, health endpoint, startup/init/teardown flow - but limited evidence of systematic retries with jitter, circuit breakers, structured tracing/correlation ids, or alerting/metrics integration.
Evidence
ruslanmusin-ai/lighter/handlers/food.py (try/except blocks with logger.exception and graceful user-facing messages) ruslanmusin-ai/lighter/api/index.py:on_startup, health (init_db usage and health endpoint) ruslanmusin-ai/songo-ai-avatar-pipeline/songo_pipeline/pipeline.py:emit_progress and manifest writing (progress events and run manifest for post-run inspection)
Expertise
Backend AI & LLM• Senior
Python• Senior
Messaging & Real-time• Middle
Industries
Media & Entertainment• Middle
Health Care• Middle
Technologies
Python• Senior
PostgreSQL
FastAPI
Asyncio
HTTPX
Uvicorn
aiogram
Flask• mentioned only
Recommendations
  • Lead development of AI-backed orchestration services - model integration, provider fallbacks, and artifact pipelines (use the songo_pipeline modules as a baseline).
  • Implement subscription and payment flows for chat-first consumer apps - webhook handling, idempotent billing writes, and lifecycle reconciliation (use lighter/api/index.py and db/database.py as templates).
  • Build content-generation and media pipelines that integrate multiple providers with robust fallbacks and local demo modes (extend live_services.py patterns).
  • Harden operational maturity - add automated tests, CI/CD, retry/backoff policies for external calls and structured tracing/metrics integration.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
lighter songo-ai-avatar-pipeline