Extending the engineering boundary.

Customer-scoped persistent memory with consent, TTL and deletion already runs in the agentic platform, where remembered text is contextual evidence and cannot authorize work. The open question is how memory behaves as it grows: which entries stay worth retrieving, how retention and compaction should be decided, and how relevance is measured rather than assumed.

Extending existing hybrid retrieval, reranking, citation-integrity and source-grounded context work by exploring how evidence should be selected, ordered and compressed before generation instead of passed through as a naive top-k chunk dump.

Exploring when explicit entity and relationship structure improves retrieval compared with vector-only RAG, especially for multi-hop and relationship-heavy questions.

Extending existing DeepEval, bilingual reranker and citation-validation work into repeatable benchmark suites that measure retrieval quality, reranking behavior, citation integrity, context efficiency, latency and failure modes.

Building a production-style local infrastructure lab with modular Terraform, the Docker provider, state migration through moved blocks, native Terraform tests, CI validation and security scanning.

Extending existing OpenTelemetry, Jaeger, Prometheus, Grafana and rollout-verification work toward AI-platform-specific signals across model serving, retrieval, agent execution and release automation.

At-least-once delivery, idempotent consumers, bounded retries and a transactional outbox are already implemented and benchmarked in the commerce platform. The open direction is deeper: reasoning explicitly about partial failure — what a service should do when a workflow fails halfway, when eventual consistency is an acceptable trade-off rather than a shortcut, and where a retry helps versus where it turns a transient failure into a duplicate or a cascading one.

Locust-driven load testing and query-plan-driven latency optimization already produced measured, reproducible numbers on the commerce platform and the ModelOps benchmark suite. The open direction is treating concurrency itself as an engineering variable: profiling to find where time is actually spent, distinguishing CPU-bound from I/O-bound from contention-bound workloads, and using Go's goroutine and channel model as a second concurrency substrate to compare against Python's async/concurrency model.

FastAPI service boundaries, PostgreSQL transactional writes and idempotency-key design are already implemented across the agent, ModelOps and commerce platforms. The open direction moves underneath the endpoint: HTTP connection and timeout behavior, gRPC as an alternative to REST for internal service communication, transaction isolation levels and connection pooling, and the operating-system behavior a production service actually runs on top of.