> forward_deployed
Forward Deployed Engineering
The role I actually do: ship into enterprise environments end to end, from cold-start discovery through architecture, implementation, and production deployment inside customer-managed clouds. Comfortable as the sole engineer on a zero-to-one build or embedded inside a customer's engineering team.
Eight capabilities, each with the techniques used and the systems they shipped in. Hover a card to flip it; click to go deep. Full track record on the experience page.
How I take an AI system from a blank whiteboard inside a customer's environment to an architecture that survives production, multi-cloud, and identity reviews.
The full retrieval stack I build from scratch, from chunking and hybrid search to GraphRAG on Neo4j, and how it cut hallucination from ~15% to under 5%.
The optimization loop after RAG works: reranking, context compression, query transformation and rerouting, and evaluation that proves each change earned its complexity.
The observability and evaluation layer I build under every LLM system: per-tenant cost telemetry, quality metrics (RAGAS, coverage scores), and dashboards stakeholders trust.
Context compression, prompt-payload reduction, and model rerouting: the token engineering that cut per-tenant LLM spend without cutting quality.
Cloud sizing, per-tenant cost telemetry, and infrastructure choices that keep an AI platform affordable in the customer's own cloud bill.
My decision framework for fine-tuning: LoRA adapters where style and format must be learned, retrieval where facts live, and evaluation gates before any tuned model ships.
Guardrails as layers, not vibes: input/output filtering, grounded-answer contracts, deterministic gates on anything that writes, and evaluation that proves the fences hold.
The operational spine under my AI systems: CI-integrated pipelines, guardrails, multi-tenant RBAC, and deployments that survive customer change windows.