Kashish Khan
NLP & LLM Engineer · MSc AI · Hackathon Winner
I build RAG pipelines, knowledge graphs, and LLM systems, from embedding search and vector retrieval to FastAPI backends and cloud deployment. MSc AI graduate seeking engineering roles in Ireland.
Deterministic intent matcher. No API key, no hallucinated answers.
Lette Inbox Copilot
GiveAGo × Lette AI · 2026
- Won the Lette AI track at GiveAGo 2026; received direct outreach from the Lette co-founder regarding open engineering roles.
- Five-module pipeline: JSON email parsing → rule-based issue classification → deterministic urgency scoring → thread-level aggregation → Streamlit UI with ranked queue, filters, and per-thread timeline.
- Hybrid design: deterministic scoring handles all triage without API dependency; GPT-4o-mini enrichment adds thread summaries and recommended actions only where needed, with graceful deterministic fallback if the LLM is unavailable.
GraphRAG Clinical Decision Support
NLP / Knowledge Graph / RAG · 2025–2026
- End-to-end five-stage pipeline over 50,000+ MIMIC-IV clinical notes: section-aware chunking (7 note types, type-specific token targets) → BioClinicalBERT embeddings (768-dim, CLS pooling) → FAISS indexing → Neo4j knowledge graph → hybrid retrieval.
- Hybrid ranking combining dense FAISS retrieval (85% weight), BM25 lexical reranking (25%), KG concept overlap with IDF-style frequency penalties (15%), and section-bias inference, catching exact drug names and ICD codes that dense models miss.
- Child-span refinement re-segments parent chunks into sub-spans and re-ranks by cosine similarity at query time, achieving sentence-level precision without storage overhead. Within-admission scoping prevents cross-patient data leakage.
- Fine-tuned BioClinicalBERT for clinical NER; entities enriched with ICD-10/ICD-9 codes and stored as Neo4j concept nodes. Evaluated across 50 tasks / 107 cases with ROUGE-L, BERTScore F1, hit@k, and precision@k.
LLM Trace & Cost Studio
Azure / Docker / FastAPI · 2025–2026
- Self-hosted LLM observability platform (LangSmith-lite): FastAPI backend, Streamlit dashboard, SQLite store, Docker Compose, deployed to Azure Container Apps.
- Captured per-request latency, token usage, and cost across LangChain pipelines; built deterministic chunk-ID mapping and embedding-drift mitigation.
Dissertation: GraphRAG for Clinical Decision Support. Modules: Advanced NLP, Deep RL, Cognitive & Ethical AI, Cloud Computing. Built MiniBioClinicalBERT in Julia from scratch.
Advanced software engineering, data structures, algorithms, and databases. Full-stack and data-driven projects in Python, Java, and SQL.