FinTech
- Agent runtime abstracted into a versioned shared library: execution loop, streaming, session lifecycle, failure semantics
- LLM gateway owning provider failover, circuit breaking, token-bucket rate control and prompt-cache boundaries
- Compliance enforced beneath the agent layer: PII masking, log redaction and audit as platform guarantees rather than agent responsibilities
- Unified retrieval interface over federated sources, served from a medallion ingestion pipeline
- Memory tiered by lifetime and models tiered by task, with per-token cost attribution across both
- Regression evaluation on a versioned golden set, judge performance measured separately from system performance
Cloud-Native AI
- Domain-adapted open-weight models for natural-language-to-kubectl translation, self-served on vLLM
- Time-series forecasting of pod resource envelopes, surfaced as vertical autoscaling recommendations
- RAG decomposed into interchangeable nodes: ingestion, indexing, retrieval, generation
Automotive
- Retrieval assistant for live vehicle consultation, deployed to production across European subsidiaries
- Market-specific manual corpora structured and metadata-filtered, safety responses grounded in source text
- Million-record vehicle dataset modelled, dealer and lead data synchronised across eight markets
- Langfuse — proposed the platform as the single source of LLM configuration; diagnosed and patched a non-ASCII serialisation bug
- Serena — orphaned dashboards now terminate on consecutive polling failure
- Graphiti — GraphRAG search-method configuration honoured; empty group_ids edge case in full-text search
- OpenClaw — internal context and tool-call text no longer surface in streamed output
Agent runtime and policy design · Agent memory architectures · Evaluation harness design · LLM observability and tracing Medium: @liebertar