Subhadip Mitra
Enterprise Data & AI Leader · Researcher
I build production multi-agent AI – and the practices and teams that scale it across the enterprise. 15+ years driving enterprise transformation across APAC as an engineering leader and AI researcher. Head of Data & Analytics, Google Cloud Southeast Asia – built the practice from zero to eight-figure cumulative delivery value, overseeing an eight-figure regional portfolio as Site Lead; 5 technical disclosures and peer-reviewed research published at ICLR 2026.
01 Experience
Progressed from Senior Consultant to Head of D&A (Year 2) to Site Lead for all SEA practices (Year 3). Member of delta – Google Cloud Consulting's innovation team delivering enterprise AI transformation from vision to production.
- Built D&A practice from zero to eight-figure delivery value across 6 countries; as Site Lead, oversaw an eight-figure regional portfolio across all 7 SEA practices – 97% CSAT and 100% revenue targets
- Architected a first-of-a-kind 15+ agent system that consolidates a global manufacturer's fragmented business-unit BigQuery projects into one self-contained foundation layer – winning a multi-million-dollar PSO and eight-figure commit
- First-of-a-kind ML notebook migration for a multinational bank via multi-agent system: 30K notebooks/4K projects, ~1 week to <30 min each, ~158K engineer-hours saved. GenAI reconciliation for a major airline; new Fortune 500 and sovereign GenAI use cases
- First-of-a-kind Enterprise Data Monetization Platform for a leading Indonesian telco – data sharing, data exchange, and sandbox-as-a-service – now the practice's blueprint
- Led delivery of a 12,000+ user analytics migration for a global shipping company
- Designed and launched JAPAC's first-of-a-kind Agent Program Office (APO), building its technical foundation – Agent SDK, agent telemetry, and attribution models measuring delivery-hour savings and efficiency gains per project and offering type – and operationalizing AI agents across 5 practices and 6 sub-regions. Made agentic assessment mandatory for all Deal Reviews, launched 7 pilot projects. Targeting 10% efficiency across a 120+ person org, extensible to the 1,400+ worldwide team
- Created Practice's Data Strategy competency and led 40+ data & AI PSO pursuits across JAPAC/SEA – exec briefings, pricing, RFPs, partner GTM; eight-figure pursuit portfolio
Led enterprise AI and data platform transformation for retail banking.
- Partnered with the Retail Bank CIO on data & AI strategy and investment decisions – business-case justification, prioritization, and multi-year planning
- Built data & analytics platform serving 11 markets, 100+ systems, and 1200+ users
- Delivered Self-Service ML Workbench with 500+ production models, reducing deployment from 6 months to 1 week
- Built credit risk models over 15K+ entities using news/social signals, materially reducing credit losses
Architected enterprise data solutions for Fortune 500 clients across APAC.
- Designed 5 data lakes processing 1.2 PB/hour and 40K daily files
- Engineered real-time platform processing 2.5M events/second, improving Ad campaign responsiveness by 80%
- Built ML fraud detection: 60% fewer false positives, multi-million-dollar savings
Founded B2B commercial vehicle marketplace (15 cities, 25+ OEM/bank partnerships). Led Thailand payments platform and bank integrations at UTU.
Windows Kernel development (Windows 7/8, Server 2012 R2), early cloud ML on Azure, CDN architecture optimization.
02 Research & open source
Quality-diversity evolution (MAP-Elites) for automatically discovering diverse LLM safety vulnerabilities. ICLR 2026 Workshop (AIWILD); extended cross-generational study on arXiv.
Speculative decoding (2-3x speedup) and custom Triton kernels (8.1x faster, 88% peak A100 bandwidth). Google Technical Disclosure pending.
Sandbagging and deception detection in LLMs via activation probes. 90-96% accuracy; steering vectors reduce sandbagging 20%. Published on PyPI.
Neuro-symbolic framework for automated distributed system generation. 274x speedup, 60% latency reduction. Google Technical Disclosure.
Runtime control of LLM agent behaviors via activation steering. 65% uncertainty detection, 100% confidence on factual queries.
Automated cloud workload discovery system for architecture assessment, built on billion-node graph models with spatiotemporal graph optimization.
03 Selected publications
Reviewer for NeurIPS and ICLR.
All publications & disclosures →04 Selected projects
Industry-agnostic agentic AI for enterprise trust decisions: a patent-pending 5-level trust cascade (rules -> ML -> adversarial multi-agent debate) with self-learning pattern system. 86% cost reduction, sub-50ms latency.
Privacy-preserving consent protocol for LLM training data. Decentralized registry with cryptographic verification and real-time opt-out enforcement.
Java 21 SMPP protocol implementation with virtual threads. 1.8M PDU decodes/sec, 1.5M encodes/sec. Published on Maven Central.