▸
Built a fully on-premise agentic AI decision-support system for LP-DED pre-deposition repair planning of flight-critical aerospace components — turbine blades, fan blades, compressor cases in Ti-6Al-4V and Inconel 718
▸
Constructed a
380-entry ChromaDB knowledge base from peer-reviewed LP-DED literature, structured across 3 layers: single-bead parameters, multi-layer deposition, and repair case studies
▸
Designed a novel
dual-query Layer 1 retrieval strategy — two parallel queries with result merging and deduplication — addressing a known RAG limitation where single queries miss entries due to varied parameter terminology across publications
▸
Implemented a
3-condition LLM grounding check before any recommendation is surfaced — ensuring suggestions are traceable to retrieved literature, not hallucinated or extrapolated beyond the knowledge base
▸
Orchestrated a LangGraph multi-agent pipeline: Retrieval Agent → Consistency Verification → Suggestion Agent+ Critic Agent → Intent Classifier(feedback Agent) — each stage independently validating before passing results downstream
▸
Implemented round-based
knowledge base expansion: engineer-approved parameter sets are written back into ChromaDB through a two-phase approval process, making the system progressively more capable through validated operational experience
▸
Deployed fully air-gapped — no cloud infrastructure, no proprietary databases — addressing aerospace traceability and IP requirements that rule out standard cloud LLM solutions
▸
Addresses a direct industry gap: existing LP-DED repair approaches are entirely operator-dependent, with knowledge lost on personnel changes and no mechanism for systematic capture — this system provides a traceable, knowledge-grounded alternative