SHIVAM CHAND KAUSHIK

• Built a four-agent LangGraph pipeline (Defect Describer, Root Cause Analyzer, Severity Classifier, Recipe Advisor) that fuses DINOv2 visual embeddings, SECS/GEM-format equipment telemetry, and a Qdrant historical defect retrieval database into a unified LLaVA-1.6 context — generating structured Failure Analysis reports in 48 seconds, replacing a 2–4 hour manual process.
• Trained a DINOv2 + MLP defect classifier on SemiFA-930, a self-constructed 930-image dataset spanning 9 semiconductor defect classes (SEM, optical, wafer map); achieved 92.1% accuracy / 0.917 Macro F1 with only 214K trainable parameters vs. 23.5M for ResNet-50 (82.9%).
• Publicly released SemiFA-930 dataset on HuggingFace; paper published on arXiv (cs.CV) — arXiv:2604.13236.

• Built an end-to-end deep learning pipeline for wafer defect pattern classification on the WM-811K dataset (811K+ wafer maps, 9 defect classes) using ResNet18 and EfficientNet-B0, with a ConvAutoEncoder for anomaly detection on unseen patterns. Achieved 97.83% overall test accuracy and 91.95% defect-class accuracy on a severely class-imbalanced dataset (103,201 normal vs 104 Near-Full samples). Improved Scratch F1 from 0.48 → 0.820 by moving from 64×64 to 128×128 resolution, adding FocalLoss(γ=2) and synthetic scratch augmentation; diagnosed critical training bugs across normalization, BatchNorm domain shift, loss double-suppression, and torch.compile checkpoint corruption.
• Trained a two-stage classifier (binary Stage 1 → 8-class Stage 2) for comparison: Stage 1 achieved 98.45% binary accuracy but error propagation reduced combined accuracy to 96.74%, confirming single-stage as the correct production model. Applied 8-view Test-Time Augmentation (+0.33% overall, stabilises low-support classes) and built a v2+v3 ensemble classifier with per-class confidence thresholding and human-review routing for safety-critical predictions.

Fine-tuned retrieval on SECS/GEM documentation with FAISS vector store. Deployed as an interactive web application for equipment and process engineers in semiconductor fabs.

• RAG pipeline ingests and chunks K&S product documentation (RapidPro, ATPP, iFlex series) into ChromaDB vector store with sentence-transformer embeddings; retrieves top-k semantically relevant passages to augment LLM context and prevent hallucination on equipment-specific queries.
• Implemented conversation memory and session history to support multi-turn diagnostic dialogues; added confidence-score filtering to route low-certainty queries to human support agents.

Used simulated bonding process data (bond force, ultrasonic power, temperature, loop height) modeled on K&S machine parameters. Classified bond pass/fail with root cause analysis and applied Bayesian optimization to identify optimal bonding parameters.

• Implemented a three-mode retrieval engine: AWS Bedrock Knowledge Base (vector similarity), BM25 keyword search, and a hybrid mode that alpha-weights and normalises scores from both — configurable at runtime via the Streamlit UI sidebar.
• Built conversation memory preserving 6-turn history and a regex-based user context extractor (name, order number, email) that personalises responses across the session; conversations exportable to JSON for audit trails.
• Deployed end-to-end on AWS: Docker image stored in ECR; Streamlit app served via ECS Fargate (serverless containers); knowledge base documents in S3; secrets managed in SSM Parameter Store; logs streamed to CloudWatch; IAM roles scoped to least-privilege Bedrock and S3 access.

• Built a GraphRAG pipeline over K&S equipment manuals and SECS/GEM standards — extracting entities (Equipment Models, Components, Alarm Codes, Failure Modes, Process Parameters) and typed relationships (TRIGGERS, CAUSES, REMEDIATED_BY, DEPENDS_ON) using spaCy NER + LLM-assisted relation extraction into a Neo4j property graph.
• Designed a hybrid retrieval engine: vector similarity (Qdrant) for dense chunk retrieval paired with Cypher-based multi-hop graph traversal — e.g. "Which bonding parameters are linked to Alarm 0x4F3 across iFlex-T4 and RapidPro?" resolved as a 3-hop Cypher query across Equipment → Alarm → Parameter nodes.
• Outperformed flat-RAG by 31% on multi-hop diagnostic queries (140 domain-expert Q&A pairs); LLM-generated Cypher queries achieved 87% structural correctness on held-out graph queries.

• Constructed a semiconductor research knowledge graph by ingesting 3,200+ papers from the Semantic Scholar API — extracting structured entities (Defect Type, Material, Detection Method, Dataset, Equipment, Metric) and typed relationships via LLM structured-output extraction (Claude API, JSON schema-constrained), yielding a Neo4j graph of 24,000+ nodes and 61,000+ relationships.
• Designed a 5-entity domain ontology grounded in semiconductor CV literature; automated entity resolution (fuzzy deduplication + embedding similarity) reduced duplicate nodes by 34%.
• Developed a LangGraph agent with Cypher tool-use: natural-language queries are schema-aware — the agent generates, validates, and executes Cypher against Neo4j, then synthesises grounded answers. Benchmarked against flat-RAG on 80 Q&A pairs: +28% answer correctness on multi-hop questions, zero hallucinated citations vs. 11% for flat-RAG.

• Built a Model Context Protocol (MCP) server exposing 7 semiconductor domain tools (Neo4j Cypher query, Qdrant semantic search, SECS/GEM alarm lookup, DINOv2 defect classifier, process parameter retrieval, SemiFA report trigger, wafer map annotator) — compatible with Claude Desktop, Cursor, and any MCP-compliant client.
• Architected four Agent-to-Agent (A2A) protocol agents — each publishing an AgentCard declaring capabilities and skills per the A2A spec: Defect Analyst, Root Cause Reasoner, Process Optimizer, and Report Composer — enabling dynamic agent discovery and task delegation without hardcoded orchestration logic.
• Demonstrated cross-agent delegation: Defect Analyst classifies a wafer image via MCP DINOv2 tool → posts an A2A task to Root Cause Reasoner → Reasoner queries Neo4j KG via MCP Cypher tool → Process Optimizer returns Bayesian-optimized parameter corrections. Full pipeline completes in under 12 seconds.
• MCP server and all four A2A agents containerised with Docker Compose; MCP compatibility verified with Claude Desktop; A2A task routing tested with Google's a2a-sdk reference client.