VP, Cloud & Platform Architecture — Deutsche Bank Building production AI/ML and GenAI systems in regulated financial services.
I architect and build production Data and AI systems — and I lead the strategy, governance, and teams that make them ship in the hardest environments. My work sits at the intersection of hands-on engineering (GenAI, NLP, data platforms on GCP) and enterprise leadership (operating models, risk frameworks, cross-functional delivery).
- GenAI platforms — LLM orchestration, guardrails, evaluation pipelines, and observability for production GenAI deployments
- RAG systems at enterprise scale — chunking strategies, hybrid retrieval, reranking, and domain-specific evaluation on financial documents
- Cloud-native data platforms — BigQuery, Dataflow, Composer, dbt, Data Vault 2.0 for regulatory-grade data foundations on GCP
- NLP & deep learning pipelines — from capsule networks and BiLSTM-CRF to transformer fine-tuning and event-driven stream processing
- AI governance & compliance — risk classification, model lifecycle controls, EU AI Act alignment, prompt audit trails
- Platform modernization — migrating legacy application estates to resilient systems on GCP
- Operating model design — team structures, delivery standards, and accountability frameworks for large engineering portfolios
DCI: Structured Collective Reasoning with Typed Epistemic Acts — arXiv:2603.11781
Introduces deliberative structure for multi-agent LLM reasoning: 4 archetypes, 14 typed epistemic acts, convergent flow algorithm. +0.95 over debate on non-routine tasks, 9.56 on hidden-profile tasks (best in study). Honest result: 62x token cost, fails on routine decisions.
The Provenance Paradox in Multi-Agent LLM Routing — arXiv:2603.18043
Self-claimed quality routing performs worse than random when delegates inflate scores. Introduces delegation contracts, claimed-vs-attested identity, and typed failure semantics for LDP. Attested routing: d=9.51, p<0.001. ldp-research
LDP: An Identity-Aware Protocol for Multi-Agent LLM Systems — arXiv:2603.08852
Introduces rich delegate identity cards, progressive payload modes, governed sessions, and trust domains for multi-agent delegation. 12x latency gains, 37% token reduction. ldp-protocol · ldp-research
| Repository | What It Does |
|---|---|
ldp-protocol |
LLM Delegate Protocol — identity-aware communication for multi-agent systems (Rust reference implementation) |
enterprise-rag-bench |
RAG patterns benchmarked for enterprise: 5 chunking strategies, 5 retrieval patterns, evaluation harness, guardrails, observability |
enterprise-genai-platform |
Reference architecture for LLM applications in banking: chain routing, prompt registry, guardrails, eval pipelines, drift monitoring |
applied-nlp-research |
Production NLP from pre-LLM to post-LLM: capsule networks, BiLSTM-CRF, transformer NER, event-driven stream processing |
ai-governance-framework |
AI governance for regulated industries: risk assessment, model lifecycle, compliance mapping, responsible AI standards |
reference-data-platform-gcp |
GCP data platform reference architecture: BigQuery, Dataflow, Composer, dbt, Data Vault 2.0 |
- Deutsche Bank — VP, Cloud & Platform Architecture (2021–present)
- Halialabs — Chief Scientist, AI & Platform Architecture (2018–2021)
- Cognizant — Technology Specialist / Software Architect (2012–2017)
- Syntel — Software Engineer / Project Lead (2007–2012)
Education: PGPMAX (Executive MBA), ISB · M.Tech Business Analytics, NUS · Google Cloud Professional Cloud Architect
sunilprakash.com — Technical writing on enterprise AI, RAG patterns, data architecture, and AI governance.
Architect + Builder + Leader · sunil@sunilprakash.com · LinkedIn