Expertise / 03
Production AI engineering and agent development
I build and improve AI functionality inside existing software products. The work is the surrounding application system: choosing and passing context, coordinating model and tool steps, checking outputs and making failures visible to the people who operate the feature.
AI features that fit the product around them
A model call needs a defined job, the right application context and a safe way to return its result. I integrate model behaviour with existing permissions, APIs and user workflows rather than treating a prompt as a standalone product.
- Claude and other supported model integrations
- Amazon Bedrock integration in AWS applications
- Structured context, customer-specific instructions and tool calling
- Output validation and human review where the use case calls for it
Orchestration, quality and operational trade-offs
Multi-stage generation and agent workflows introduce dependencies that need engineering attention. Changes to instructions or context can affect later steps, while provider latency, cost and partial failures affect the user experience. I help make these behaviours observable and establish examples and checks for evaluating changes. The application still needs dependable API integrations for tools and data providers, and appropriate AWS infrastructure for deployment and operation.
- Agent orchestration using Strands Agents
- Multi-stage LLM workflows and external tools
- Prompt and context iteration based on observed output issues
- Evaluation examples, failure handling and fallback behaviour
Production experience, described accurately
At HoneySales, I work alongside the CTO on AI capabilities in a production B2B SaaS platform. Company research and enrichment provide context for LLM-powered email and LinkedIn message generation, including customer-specific instructions and multi-stage processing. The system uses Claude Haiku and Sonnet; earlier work also included Amazon Nova. This is ongoing product engineering, and output quality continues to require review and improvement.
Relevant professional experience
This work concerns application engineering and model integration. I do not train foundation models or claim that probabilistic output can be made perfectly deterministic.
Questions buyers usually ask
Can you add LLM functionality to an existing application?
Yes. The work can start with an existing user journey and define where model output, application data and human decisions belong.
Do you train custom foundation models?
No. My focus is integrating supported models into production applications and engineering the context, orchestration, evaluation and operational behaviour around them.
Discuss an engineering engagement
Tell me about the work.
Share the system, the technical challenge and the kind of commitment you are considering.
Contact Serhii ↗