/projects
Selected work.
Most of this is proprietary, so descriptions are deliberately light on internal detail. Open-source side projects live on GitHub; peer-reviewed work is on the resume.
-
Enterprise RAG System
Production RAG over 500K+ documents — hybrid search (vector + BM25), HyDE rewriting, cross-encoder re-ranking, RAGAS evaluation. 92% recall@10, sub-3s responses.
-
Agentic AI Platform
Multi-agent orchestration with LangGraph + CrewAI (ReAct, Plan-and-Execute) across 25+ tools, with episodic and knowledge-graph memory, LangSmith tracing. 99.2% task completion.
-
LLM Guardrail System
Multi-layer guardrails (Guardrails AI + NeMo + custom) with PII detection, SHAP explainability, monitoring app (React / FastAPI / PostgreSQL). 89.2% blocked, 5K+ concurrent.
-
Multi-Tenant NER Platform
Scalable NER serving multiple enterprise clients with tenant-specific isolation and SLA-based model selection. Sub-200ms inference.
-
Code-Generation Fine-tuning
LoRA fine-tuning of large code models on Hugging Face for domain-specific completion. +40% accuracy, −90% training cost, A/B adapter swapping.
-
ISO 20022 / SWIFT NLP Pipeline
NLP pipelines in Python to validate and auto-generate compliant financial messages (ISO 20022, SWIFT MT/MX) for regulated workflows.
-
Graph-Based Fraud Detection
Graph systems on Neo4j + GraphQL over large financial datasets — link prediction, clustering, anomaly detection, with Pandas/Spark mining for real-time signals.
-
Banking-Platform Microservices
Cloud-native banking platform: Golang for performance-critical services, Python / Django for AI orchestration. gRPC + REST, KrakenD gateway, Keycloak RBAC/ABAC, custom Kubernetes operators.
-
Healthcare Activity Recognition
Edge-to-cloud sensor-fusion pipeline (ESP-32 + Raspberry Pi). PCA / FFT / LDA preprocessing, ensemble of classical ML + CNN models, on-device mobile inference.