Michael Korostelev korostelevm@gmail.com | 267-265-3950 Philadelphia, PA https://mikekorostelev.com | https://github.com/korostelevm | https://www.linkedin.com/in/michaelkorostelev PROFESSIONAL SUMMARY Staff AI Engineer, Agent Infrastructure & RL Environments Specialize in agent infrastructure, distributed execution, and RL environments. At Pareto AI, I lead a team of 7 building the system that produces RL environments and verifiable tasks for frontier AI labs, sustaining 1,000+ concurrent solver rollouts and 100M+ tokens per minute and producing tasks 4x cheaper than human experts. Previously built durable multi-agent orchestration on Temporal at MFD.AI and co-founded Cyclic, a serverless developer platform with 150k+ signups and 1M+ deployments. ~15 years across production AI systems, distributed infrastructure, and hands-on leadership of teams up to 20 engineers. SKILLS Hands On Technical Skills: AI/Agent Systems: RL environments, Temporal.io, Claude Agent SDK, OpenAI/Claude APIs, multi-turn tool use, sandboxed code execution, eval frameworks, Languages: Python, TypeScript, Node.js, Rust, Bash, Infrastructure: AWS (ECS, EC2, Lambda, S3, RDS, Step Functions, SageMaker), GCP (Vertex AI), Terraform, Docker, Databases: PostgreSQL, MongoDB, Redis, DynamoDB, Architecture: Distributed workflows, event-sourced replay, event-driven systems, microservices, serverless, CI/CD & Observability: GitHub Actions, Datadog, CloudWatch, Honeycomb, Sentry, LangSmith, Frontend: React, Next.js, Vue, TailwindCSS EXPERIENCE PARETO AI - Staff AI Engineer Remote, US | 12/25 - Present Lead a 7-person Applied AI team building the system that produces RL environments and verifiable training tasks for AI labs, using agents to generate tasks and calibrate their difficulty against solver models. • Architected a system that produces RL tasks and environments. Teams of agents write and revise tasks in Docker sandboxes while solver-model rollouts measure each task's pass rate and other metrics, iterating until the task reaches a target difficulty range to maximize learning rate. Sustaining 1,000+ concurrent solver rollouts and 100M+ tokens per minute alongside hundreds of concurrent authoring agents. • System produces or calibrates tasks 4x cheaper than human experts for the same tasks, validated by domain experts and by client QA at two major AI labs. • Co-authored "Trading Human Curation for Synthetic Augmentation in RLVR" (arxiv.org/abs/2606.03800, 2026), which measures how well synthetic tasks substitute for human-authored RLVR tasks; they retained held-out generalization across a ten-benchmark suite spanning code, reasoning, instruction following, and agentic function calling. • Designed and built a distributed agent framework on Temporal for long-running agents. Adapted multiple providers/harnesses (Claude Code SDK, OpenCode, OpenAI) to operate in a distributed environment, with checkpoints that can be resumed or forked and persistent workspaces, so agents survive worker restarts and release compute while waiting on deferred tools and sub-agents to spread a single agent's work across a heterogeneous network of workers. • Set agentic engineering standards and automated guardrails, giving agentic reviewers production awareness to verify scenarios and run smoke tests automatically • Tech Stack: Python, FastAPI, Temporal.io, PostgreSQL, Terraform, AWS (ECS, EC2, S3, RDS, Bedrock), GCP (Vertex AI), Docker, Claude Agent SDK, Claude/OpenAI APIs, Datadog, React MFD.AI - Lead AI Engineer Los Angeles, CA / Remote, US | 11/24 - 12/25 Architected and built an AI orchestration platform automating sourcing, finance, and operations through multi-agent workflows at a supply chain startup. • Built multi-agent orchestration on Temporal where every LLM response and tool result is recorded in event history, so long-running agent workflows can be replayed, resumed after failures, and audited. • Made agents, workflows, and tools interchangeable, so any workflow can call an agent or another workflow as a tool, and added checkpointing and exact state restoration for long-lived processes. • Added controls for predictable agent behavior (side-effect isolation, tool permissions, loop protection, context pruning) and human-in-the-loop steps via Temporal signals, with Slack escalation and a React workflow editor for live monitoring. • Scaled BDR capacity 10x (50 to 500 monthly conversations per rep) and cut vendor quote turnaround from 30 days to 1 week by automating outreach, follow-ups, CRM sync, and approval routing across 100+ tools via an MCP-based tool layer. • Tech Stack: Temporal.io, Node.js, TypeScript, MongoDB, AWS (Lambda, ECS, S3), Terraform, Docker, LangSmith, React SIGNANT HEALTH - Senior Manager, Engineering Blue Bell, PA / Remote, US | 1/24 - 11/24 • Managed three remote engineering teams (15 engineers across the US, Romania, and Finland) delivering core infrastructure for regulated clinical-trial SaaS products. • Led development of an integrated developer platform automating CI/CD, environment provisioning, and compliance-driven deployments across teams, reducing release friction and audit preparation time. • Introduced LLM-generated release notes, change logs, and validation documentation in a high-compliance environment. CYCLIC.SH (CYCLIC SOFTWARE) - Co-Founder Philadelphia, PA / Remote, US | 9/21 - 12/23 • Co-founded and built cyclic.sh, a serverless PaaS for Node, Python, and Go apps; grew it from launch to 150k+ signups, 1M+ deployments, 40k+ monthly active hosted applications, and 3k+ daily active users. • Architected the platform on AWS serverless, active-active across 3 regions, deploying projects from version control to 18+ regions with user-facing dashboards for CI/CD, logs, and monitoring. • Wrote a custom AWS Lambda runtime in Rust that brought cold starts down to ~200ms. • Automated billing, CRM, and fraud detection and mitigation for tens of thousands of free-tier users. • Tech Stack: AWS (Lambda, API Gateway, DynamoDB, S3, Step Functions, CloudFormation, SQS/SNS), Node.js, Rust, Python, Vue, Docker BLACK KNIGHT FINANCIAL SERVICES - Application Architect IV, AIVA Product / Applications Development Lead Philadelphia, PA / Remote, US | 12/18 - 10/21 • Directed architecture and delivery of an AI document-processing platform handling 1M+ pages of mortgage data daily for top U.S. banks, leading an R&D team of ~20 engineers and scientists. • Moved AI inference to serverless functions, cutting cloud spend by ~40% while increasing throughput and reliability. • Built an MLOps and data-governance framework for compliant model training on PII under financial regulations. • Led SOC 2 and PCI compliance for the product, including enterprise SSO, automated audit logging, and pilot-light disaster recovery with 15-minute RPO/RTO. • Promoted from Senior Engineer to Application Architect within three years. • Tech Stack: AWS (Lambda, Step Functions, SageMaker, Textract, DynamoDB, S3, Glue, Athena), Node.js, Python, Docker, CDK LEGAL SCIENCE LLC. - Founding Engineer Philadelphia, PA | 5/15 - 12/18 • Principal investigator on a 2018 NSF SBIR Phase I grant, "Machine Assisted Comparative Policy Analysis in Public Health." • First engineer at a legal tech startup; built the annotation software legal researchers used to code legislation for quantitative and machine-learning analysis. • Built supervised ML models to classify legal text, trained on data collected with the annotation tool. TEMPLE UNIVERSITY - Adjunct Faculty, Research Assistant Philadelphia, PA | 5/09 - 5/15 • Developed and taught Engineering Design Principles (Tyler School of Art & Architecture) and Digital World, a gen-ed course on digital technology and networking. • Research and software development in computer vision, multi-view camera calibration, and simulation for aerial target tracking. • Program Mentor at Apps & Maps Design Studio, a youth technology incubator; managed project teams and wrote grant proposals, with projects covered in print and television media. • Consulting on startup MVP/POC projects, including • automated patient health monitoring via SMS with Fox Chase Cancer Center • hardware and web dashboard for a child-tracking GPS module with Brams Inc. • posture and stability data capture with Temple College of Public Health EDUCATION Ph.D. Program, ABD (all but dissertation), Engineering Temple University | Philadelphia, PA | 2011 - 2015 Master of Science (MS), Electrical and Computer Engineering Temple University | Philadelphia, PA | 2011 Thesis: Multi-view projector-camera calibration for computer vision and augmented reality applications Bachelor of Science (BS), Electrical and Computer Engineering Temple University | Philadelphia, PA | 2009