Subhadip Mitra
Applied AI & Data · Engineering Leader · AI Researcher
I build AI systems that move from research to production at enterprise scale. 15+ years across APAC, currently Head of Data & Analytics for Google Cloud Southeast Asia and Site Lead for the regional professional services organization, and still writing the systems underneath: multi-agent platforms, evaluation infrastructure and GPU inference kernels.
01 Experience
Progressed from Senior Consultant to Head of D&A (Year 2) to Site Lead for all SEA practices (Year 3).
- Agent Program Office. Conceived and launched a JAPAC Agent Program Office for a 700+ person organization: secured executive sponsorship, defined the pilot, governance and efficiency-measurement model, and personally built the Agent SDK and telemetry backbone. Positioned for potential expansion to a 2,000+ person global organization.
- Agent system in production. Built a 20+ agent system that reconstructs the lineage graph of a dbt/BigQuery estate 25+ layers deep, breaks circular dependencies, and generates and validates every restructured model itself. Rebuilt the most complex business unit in six hours. Technical core of an account recovery.
- Migration accelerator. Built a four-agent migration accelerator, in production at a leading Asian bank: ~30K notebooks, ~4K ML projects, modeled effort down 97.5% (160K to 4K hours). A representative project moved from about a week to under 30 minutes.
- Practice and site leadership. Built the Data & Analytics practice from zero to eight-figure cumulative delivery value across six countries. As Site Lead, govern delivery across the region's AI/ML, Data, Infrastructure, Application Modernization and Security practices.
- Pursuit portfolio. Led and shaped the strategic Data & AI pursuit portfolio across JAPAC with sales and GTM teams, contributing to major competitive services wins including platform takeouts involving AWS, Azure, Databricks and Snowflake, and to substantially larger downstream cloud consumption. Pursuit leadership recently extended across all of India.
- Sovereign AI. # "Air-gapped" is user-confirmed (16 Aug 2026): the deployment is on # air-gapped GDC. It is the harder and more specific claim. Developing an MCP-based custom connector for Gemini Enterprise on air-gapped Google Distributed Cloud in a sovereign environment.
- Forward-deployed intervention. Selected by VP leadership for an eight-week onsite intervention on a deeply red financial-services migration; helped steer the program back to green.
- Capability building. Founded the practice's data-strategy group, competency pods, a consultant bootcamp and a community of practice; contributed to an SME data and analytics academy and a graduate traineeship with Singapore's IMDA. Hired and helped hire 10+ across India, Greater China, Japan and ANZ.
- Applied AI delivery. Sponsored by the customer's CFO and built alongside him: a production AI-assisted reconciliation platform for OTA invoices at a major airline and digital-travel business, with a custom rules engine covering 2,000+ invoice formats across PDF, spreadsheet and image inputs, multi-language and multi-currency. ~97% straight-through processing, with 10K+ line items reconciled in under 25 minutes against a manual baseline of days.
- Executive advisory. Advise CIOs, CTOs, CDOs, CAIOs and CFOs, present in Google Executive Briefing Centers, and partner with GTM directors and regional consulting leadership on account strategy.
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 a data & analytics platform serving 11 markets, 100+ systems and 1,200+ users
- Delivered a self-service ML workbench running 500+ production models, cutting deployment from about six months to about a week
- Built alternative-data credit risk models using news/social signals across 15K+ entities
Architected enterprise data solutions for large enterprises across APAC.
- Designed 5 data lakes with ETL pipelines running up to 1.2 PB/hour and 40K daily files
- Engineered a real-time platform processing 2.5M events/second
- Built ML fraud detection reducing false positives by ~60%
Founded a B2B commercial-vehicle marketplace across 15 cities with 25+ OEM and bank partnerships; built the recommender systems, vehicle auctioning platform and product analytics behind it.
Led payments-platform and bank-integration engineering across Southeast Asia.
Windows kernel components for Windows 7/8 and Server 2012 R2; early Azure ML and CDN architecture optimization.
02 Research & open source
Compiler from REST (OpenAPI, Swagger) and SOAP (WSDL 1.1) specifications to governance-aware MCP tool surfaces: field-level provenance, human approval gates for destructive operations, refusals compiled into the generated artifact. Pre-registered evaluation against deterministic state oracles. Published on PyPI.
Enterprise-scale LLM evaluation on Apache Spark: bootstrap confidence intervals, paired significance tests, multi-provider support. Published on PyPI; paper on arXiv.
Quality-diversity evolution (MAP-Elites) for automatically discovering diverse LLM safety vulnerabilities. ICLR 2026 Workshop (AIWILD); extended cross-generational study on arXiv.
Sandbagging and deception detection in LLMs via activation probes: 90-96% accuracy in sandbagging and deception-detection settings. Published on PyPI.
Industry-agnostic agentic AI for enterprise trust decisions: a tiered cascade of deterministic rules, ML and adversarial multi-agent reasoning with cost-aware routing. JAPAC Champion and Global #2, Google Cloud Professional Services Hackathon 2025, with two associated technical disclosures.
Runtime control of LLM agent behaviors via activation steering, without retraining: refusal raised to 95%+ across Mistral-7B, Gemma-2-9B and Qwen3-8B at a 5% false-positive rate, where prompting reaches full refusal only at a 100% false-positive rate. LangChain integration.
Custom Triton kernels: RMSNorm 8.1x faster at 88% of peak memory bandwidth on A100, and a portable W4A16 4-bit GEMM that beats cuBLAS FP16 by 1.1-1.3x in decode workloads, cross-vendor on A100 and MI300X. Speculative decoding at 2-3x in its benchmark setting. Two kernels published to the Hugging Face Kernel Hub.
Neuro-symbolic framework for automated distributed-system generation. 274x speedup and 60% latency reduction in its benchmark setting. Google Technical Disclosure.
03 Selected publications
Reviewer for NeurIPS and ICLR.
All publications & disclosures →04 Selected projects
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.