01
THINK ARCHITECTURALLY
Understand components, interfaces, constraints, failure modes, deployment, observability, and cost as one system.
01 / PROFILEENGINEERING-ORIENTED BUILDER
ALEXEY
PAVLINOV
Principal AI & Software Engineer
Production LLM systems · RAG · Agents · Information Retrieval · Cloud · High-throughput backends
Production LLM systems · RAG · Agents · Information Retrieval · Cloud · High-throughput backends
I design systems from first principles, prototype quickly, measure what actually works, and turn the useful parts into reliable production software. My strongest work sits where AI meets real engineering: architecture, APIs, data pipelines, retrieval, deployment, observability, performance, and evaluation.
VECTOR FIELD 06
COHERENCE 0.977
MODE / SYNTHESIS
COHERENCE 0.977
MODE / SYNTHESIS
ARCHITECTURE
SYSTEM BOUNDARIES
FAILURE MODES / COST
SYSTEM BOUNDARIES
FAILURE MODES / COST
EVALUATION
MEASURE QUALITY
BEFORE CLAIMING IT
MEASURE QUALITY
BEFORE CLAIMING IT
DELIVERY
PROTOTYPE → PROD
END TO END
PROTOTYPE → PROD
END TO END
SOFTWARE ENGINEERING
15+ Y
SYSTEMS / BACKEND / CLOUD
PYTHON
13+ Y
ASYNC · APIs · DATA · AI
STREAMING THROUGHPUT
200K/s
500K/s PEAK
ENGINEERING STYLE
E2E
REQUIREMENTS → ACCEPTANCE
CURRENT VECTOR
AI
LLM · RAG · AGENTS · IR
02 OPERATING MODEL
HOW I APPROACH TECHNICAL PROBLEMSSimple answer first.
Architecture second.
Complexity only when earned.
I like technically difficult systems, but complexity is not the goal. I want the smallest architecture that proves the idea, a way to measure whether it works, and a clear path from prototype to production. I care about precise terminology because names define boundaries, ownership, and how a team reasons about a system.
02
BUILD SOMETHING REAL
Prefer a working prototype, benchmark, command, diagram, or experiment over a vague architectural claim.
03
EVALUATE, DON'T ASSUME
Chunk size, quantization, retrieval strategy, model choice and agent topology are parameters to test—not folklore.
04
OPTIMIZE FOR VALUE
Performance, reliability and elegance matter, but so do delivery time, cost, maintainability, and business usefulness.
03 SELECTED SYSTEMS
GREENFIELD / PRODUCTION / HIGH-THROUGHPUTSYSTEM / 01 · AI AGENTS
Enterprise AI agents and RAG applications
Multi-agent systems that query enterprise sources, knowledge bases, APIs, search systems and operational tools. The interesting part is not the chat UI—it is orchestration, retrieval, tool contracts, security boundaries, evaluation and observability.
SYSTEM / 02 · MARKET DATA
Liquidity-provider streaming pipeline
Ingestion, normalization, processing and distribution for stock-exchange feeds with sustained high throughput and peaks around half a million events per second. Tooling and observability were treated as part of the system, not as afterthoughts.
SYSTEM / 03 · PROCUREMENT AI
Tender / proposal analytical system
Built from scratch to estimate proposal quality against tender requirements using an industrial knowledge base, RAG, LLMs and document intelligence. Work included system design, integration, technical strategy and end-to-end delivery.
04 ARCHITECTURE FIELD
A TYPICAL PRODUCTION AI SYSTEM / NOT A PRODUCT DIAGRAMROUTE / LIVE
DATA / RETRIEVALCONTROL / ROUTINGTOOLS / EXTERNAL SYSTEMS
05 CAPABILITY MATRIX
RELATIVE DEPTH / CURRENT EMPHASISDEPTHEXTRA BOOSTSOFTWARE / SYSTEM ARCHITECTUREPRIMARY
Backend architecture, service boundaries, APIs, async systems, reliability, technical strategy.
LLM / RAG / INFORMATION RETRIEVALPRIMARY
Embeddings, hybrid retrieval, chunking, knowledge bases, evaluation, context engineering.
AGENTS / ORCHESTRATIONACTIVE
Tool use, multi-agent topologies, graphs, workflows, MCP/A2A, memory and guardrails.
CLOUD / DELIVERYSTRONG
AWS, containers, Terraform, CI/CD, monitoring, production integration and cost-aware deployment.
DATA / STREAMING / SEARCHSTRONG
High-throughput pipelines, SQL/NoSQL, OpenSearch/Elasticsearch, vector databases and event processing.
PROTOTYPING / GREENFIELDCORE MODE
From ambiguous problem to architecture, proof, benchmark, implementation and acceptance.
06 EXPERIENCE LEDGER
SELECTED RECENT WORK2025 → NOW
REMOTE / PORTUGAL
REMOTE / PORTUGAL
Senior Software Engineer / FiveBlueSoftware
AI agents across medical, financial and e-government projects; RAG applications and enterprise integrations. Also high-throughput market-data feed systems with streaming, normalization, distribution, testing tools and operational dashboards.
AI + STREAMING
LANGGRAPH
OPENSEARCH
PYTHON
LANGGRAPH
OPENSEARCH
PYTHON
2023 → 2024
REMOTE / PORTUGAL
REMOTE / PORTUGAL
Principal Software Engineer / Softengi
Designed and delivered procurement AI prototypes and analytical systems using RAG, LLMs, vector search, knowledge bases and speech processing. Worked across technical strategy, product requirements, integration and delivery.
RAG / LLM
AWS
MILVUS
FASTAPI
AWS
MILVUS
FASTAPI
2021 → 2023
REMOTE
REMOTE
Software Engineer / The Product Engine
Backend engineering for Comcast cable-TV streaming API aggregation, including scalable cloud components, security, documentation, testing and production delivery.
DISTRIBUTED BACKEND
AWS LAMBDA
DYNAMODB
KINESIS
AWS LAMBDA
DYNAMODB
KINESIS
2018 → 2021
Software Engineer / GeeksForLess
Admin/CMS application supporting content production and game operations across multiple mobile titles, with Python/Flask services, web UI, AWS and multiple data stores.
WEB PLATFORM
FLASK
POSTGRESQL
MONGODB
FLASK
POSTGRESQL
MONGODB
07 DECISION PATTERNS
HOW I BUILD MENTAL MODELSCOMPARE / 01
Graph vs Workflow
Topology and routing control versus task ordering and dependency management. The useful question is not which is “better,” but where each abstraction gives more leverage.
BOUNDARY > BRAND
COMPARE / 02
MCP vs A2A
Agent-to-tool interoperability versus agent-to-agent collaboration across systems. Different protocol boundaries, different failure and trust models.
INTERFACE > HYPE
COMPARE / 03
Speed vs Quality
Whether testing quantization, retrieval or architecture, performance only matters alongside answer quality, correctness, operational cost and reproducibility.
MEASURE > ASSUME
08 OPEN CHANNEL
SYSTEM DESIGN / AI ENGINEERING / GREENFIELD BUILDBring me the problem
that has too many layers.
I am most useful when the task crosses boundaries: backend + AI, retrieval + evaluation, architecture + deployment, streaming + observability, or a new system that still needs its shape discovered.
ROLE VECTORPRINCIPAL / SENIOR AI ENGINEER
WORK MODEREMOTE / PORTUGAL
BEST FITGREENFIELD + COMPLEX SYSTEMS
DEFAULT QUESTIONHOW DO WE PROVE IT WORKS?