Separate candidate retrieval from reranking
Multilingual-e5 embeddings narrow the registry to plausible candidates. Claude then reranks those candidates before they enter the agent context, balancing semantic recall with context quality.
← Back to selected work
Enterprise AI Platform
An internal platform for moving enterprise AI initiatives from intake to development and delivery, with reusable agent capabilities at its core.
RoleDeveloper in a four-person core team
FocusPython · TypeScript · AWS Bedrock AgentCore
01Context
The problem
AI initiatives need a common path from an initial idea to a working delivery. Without reusable capabilities and context selection, teams repeat integration work and agents receive more context than they can use effectively.
My contribution
I co-designed the platform and implemented a large part of the solution across backend services, agent infrastructure, MCP integrations, skills, technical documentation, proposals, estimates, and demonstrations. The work is presented at the level of the team platform while separating my hands-on contribution from the wider group.
02System view
How the pieces connect
A conceptual view of the system, recreated from the public-safe project description.
01InitiativeIdea, requirements, delivery path
02CapabilitiesMCP integrations and tools
03Retrievemultilingual-e5 candidates
04RerankClaude selects relevance
05ContextSkills available to the agent
03Outcome
The platform provides cloud development for non-technical users, local workflows for technical users, and a reusable layer of MCP integrations, tools, and skills. It is built for production use; internal adoption and satisfaction figures are not published here.
7MCP integrations in the capability layer
150+tools made reusable
40agent skills in the registry
04Engineering notes
Reusable capability layer
MCP integrations, tools, and skills are treated as shared building blocks instead of one-off additions to individual agents. That supports different user paths without duplicating the entire platform.
Keep future architecture explicit
Sub-agent and A2A patterns have been evaluated for future cross-department collaboration. A2A remains under evaluation and is not represented as a deployed capability.
05Scope & limits
Stack
What I would keep visible
- This case uses a recreated conceptual diagram and does not expose internal screenshots, source code, or proprietary architecture.
- CloudWatch is the confirmed observability layer; no additional specialized observability is claimed.
- Internal percentages for satisfaction, speed, or initiative coverage are excluded because their calculation method is not validated for public use.
Context management is a product and architecture problem. Candidate retrieval followed by reranking makes skill selection intentional and helps preserve the agent's useful context window.
Explore another case