cv
Note: This is a public version with certain details removed for privacy. For a comprehensive resume including specific project metrics and contact details, please reach out via email or LinkedIn.
Senior Engineering Leader with 15+ years of experience bridging fundamental AI research and enterprise-scale system delivery. Currently leading Google Cloud’s Data & Analytics practice for Southeast Asia while driving internal innovations on LLM inference efficiency and multi-agent systems.
Proven track record of operating as a "Player-Coach": managing regional engineering portfolios while simultaneously architecting and patenting novel frameworks (UPIR, ARTEMIS, FTCS, Speculative Decoding).
Work Experience
Data & Analytics Manager | Site Lead, PSO Southeast Asia
Dual-track role combining technical innovation leadership with regional delivery management. Built Google Cloud's Data Analytics practice across Southeast Asia while serving as Site Lead overseeing cross-practice operations. Member of delta - Google Cloud's innovation and transformation team architecting enterprise AI solutions at scale.
Strategic Leadership & Delivery
- Practice Leadership: Built Data Analytics practice for Southeast Asia from 0 to 1, establishing the region's premier capability serving 7 countries with strategic enterprise clients.
- Regional Operations: Serve as Site Lead overseeing PSO Southeast Asia delivery operations, directing cross-functional teams while maintaining 97% CSAT and contributing to 100% annual revenue target attainment.
- Portfolio Management: Direct \$ XXM+ Data Analytics delivery portfolio across JAPAC while simultaneously overseeing \$ XXM+ cross-practice portfolio as regional Site Lead.
- Strategic Interventions: Led critical engagements for JAPAC strategic accounts including major financial services institutions and consumer electronics manufacturers, ensuring delivery excellence and client success.
- Enterprise Delivery: Executed high-impact projects including 12K+ user analytics migrations, first Data & AI Centers of Excellence, Data Monetization Platforms, and petabyte-scale data platform modernizations.
- Executive Advisory: Partner with C-level stakeholders (CTOs, CDOs) to define data modernization and AI transformation roadmaps, translating technical capabilities into business outcomes.
- Agentic AI Transformation: Pioneered organization-wide adoption of agentic AI across PSO JAPAC for both customer solutions and internal productivity, including autonomous data engineering agents.
Technical Innovation & Research (Official IP)
- LLM Inference Efficiency: Research on speculative decoding, custom Triton kernels, and KV-cache compression strategies. Filed Google Technical Disclosure on hybrid compression systems for multi-tenant serving optimization.
- Distributed Systems Synthesis (UPIR): Invented neuro-symbolic framework combining formal verification and reinforcement learning to automate distributed system generation - achieved 274x speedup in synthesis with 60% latency reduction.
- Context Architecture (FTCS): Designed Field-Theoretic Context System modeling context as continuous fields to address long-horizon memory fragmentation in AI agents. Published as Google Technical Disclosure.
- Data Processing for GenAI (ETLC): Authored whitepaper introducing Extract, Transform, Load, Contextualize framework adding semantic, relational, and behavioral context to data pipelines for RAG and agentic systems.
- Multi-Agent Framework (ARTEMIS): Created adaptive debate-driven decision framework for enterprise multi-agent systems. Published as Google Technical Disclosure.
- Intelligent Trust Engine (CatchMe): Developed industry-agnostic agentic AI system for enterprise-scale trust decisions across Finance, Healthcare, Insurance, Cybersecurity, and Supply Chain. Features APLS (self-learning pattern synthesis) and five-level cascade routing achieving 86% cost reduction with sub-50ms latency. Won Google Cloud PSO Hackathon JAPAC Regionals, qualified for World Finals. Two pending Google Technical Disclosures.
Principal Engineer - Data & Analytics Transformation
Led enterprise-wide AI and data platform development serving 11 markets and 1200+ global users, delivering technical excellence while influencing C-suite data strategy.
- Delivered a Self-Service ML Platform that reduced model development time from 6 months to 1 week
- Designed credit risk AI models integrating alternative data sources, improving accuracy by 15%
- Modernized MarTech infrastructure, driving 30% increase in customer acquisition
Principal Data Engineer / Solution Architect
Architected enterprise-scale data solutions for Fortune 500 clients across APAC, designing scalable platforms with measurable business impact.
- Engineered 5 high-performance data lakes processing 1.2 PB/hour, achieving 20% optimization
- Built real-time fraud detection systems, reducing false positives by 60% and saving $XM annually
- Designed enterprise architectures supporting global Fortune 500 clients across APAC
Software Engineering, Architecture and Technical Consulting Roles
Progressively advanced through roles in software development, systems integration, and technical consulting within financial services and algorithmic trading domains.
Research & Open Source Engineering
LLM Inference Efficiency Research
Research implementations addressing the fundamental bottleneck in LLM inference: memory-bandwidth constraints rather than compute limits. Explores acceleration through speculative decoding, custom GPU kernels, and quantization strategies.
- Speculative Decoding Suite: Six techniques including standard speculation, tree speculation, EAGLE-style drafting, Medusa multi-head, KV-cache compression (8x compression via INT8/INT4 quantization + H2O eviction), and diffusion efficiency optimizations. Production systems report 2-3x inference speedup.
- Custom Triton Kernels: High-performance GPU kernels for transformer operations - RMSNorm (8.1x faster, 88% peak bandwidth), fused RMSNorm+Residual (6.0x speedup), SwiGLU (1.6x), INT8 GEMM (2x memory savings). Demonstrates memory-bandwidth optimization from 11% to 88% of A100 peak.
- Device-Agnostic Implementation: Supports CUDA, Apple Silicon (MPS), and CPU with full KV-cache integration reducing complexity from O(seq²) to O(seq).
AI Metacognition Toolkit
Activation-level detection of sandbagging, deception, and situational awareness in LLMs. Linear probes achieve 90-96% accuracy across Mistral, Gemma, and Qwen models. Includes steering vectors for runtime behavior control.
- Sandbagging Detection: Linear probes trained on activation differences detect sandbagging intent with 90-96% accuracy. Model-specific representations - no cross-model transfer.
- Steering Vectors: Activation steering reduces sandbagging behavior by 20% in Gemma models without retraining.
- Bayesian Situational Awareness: KL-divergence based detection of behavioral changes and "Observer Effects" during interaction.
- Engineering Rigor: 275 test cases, 95% code coverage, type-safe implementation, published on PyPI.
UPIR - Distributed Systems Synthesis
Neuro-symbolic framework combining formal verification, program synthesis, and reinforcement learning to automatically generate verified distributed system implementations from specifications.
- Compositional Verification: SMT-based verification engine with proof caching achieving 274x speedup for 64-component systems.
- CEGIS Synthesis: Counterexample-guided inductive synthesis with constrained PPO optimization preserving formal guarantees.
- Performance: 60.1% latency reduction, 194.5% throughput increase, 89.9% pattern reuse potential in benchmark tests.
Steering Vectors for Agent Behavior Control
Runtime control of LLM agent behaviors through activation steering vectors - modifying model outputs at inference time without retraining. Demonstrates more calibrated control than traditional prompting approaches with LangChain integration.
- Contrastive Activation Addition: Extract steering vectors from contrast pairs and inject into model activations for behavior modification.
- Uncertainty Calibration: Achieves 65% uncertainty detection on ambiguous questions while maintaining 100% confidence on factual ones - superior to prompting which causes indiscriminate hedging.
- Multi-Vector Composition: Dynamic strength adjustment per-request with interference mitigation for combining multiple behavioral controls.
- Production Ready: LangChain integration, tested on Mistral-7B, Gemma-2-9B, and Qwen3-8B models.
Spark LLM Eval - Distributed Evaluation Framework
Distributed LLM evaluation framework built on Apache Spark for enterprise-scale model assessment. Addresses the gap in evaluating LLMs at scale with statistical rigor, integrating seamlessly with Databricks infrastructure.
- Distributed Processing: Pandas UDFs with Arrow for efficient batching, scales linearly across Spark executors for millions of examples.
- Statistical Rigor: Bootstrap confidence intervals, paired significance tests (t-tests, McNemar's, Wilcoxon signed-rank), and effect size calculations.
- Multi-Provider Support: Works with OpenAI, Anthropic Claude, Google Gemini, and vLLM with smart rate limiting (token bucket algorithms).
- Enterprise Integration: MLflow experiment tracking, Delta Lake versioning, and comprehensive metrics (lexical, semantic, LLM-as-judge).
CatchMe - Intelligent Trust Engine
Industry-agnostic agentic AI system for enterprise-scale trust decisions across Finance, Healthcare, Insurance, Cybersecurity, and Supply Chain. Uses adversarial debate protocols (prosecutor/defense/judge) to filter hallucinations and build audit trails for regulated environments.
- APLS (Automated Pattern Learning System): Self-learning system that observes expensive AI decisions and automatically synthesizes cheaper deterministic rules over time. Achieves 76% cost reduction after 6 months of pattern learning.
- Five-Level Cascade Routing: Intelligent transaction routing based on cost-benefit analysis. Automatically escalates high-value/high-risk decisions to AI reasoning while routing routine cases through deterministic rules (e.g., $50K wire with 75% confidence → AI escalation with ROI 1,250,000x). Sub-50ms latency for most transactions.
- Multi-Agent Consensus: Adversarial debate system where agents act as prosecutor, defense, and judge to ensure decision quality and create compliance audit trails for regulated industries.
- Production Impact: 86% cost reduction vs AI-only approaches ($20.7K vs $150K for 10M monthly transactions), serving Finance (fraud detection), Healthcare (claims processing), Insurance (policy validation), Cybersecurity (anomaly detection), and Supply Chain (verification).
- Recognition: Winner - Google Cloud PSO Hackathon JAPAC Regionals, Qualified for World Finals. Built on Vertex AI, Gemini 2.x, BigQuery ML, Cloud Run.
Education
Publications & Technical Disclosures
UPIR: Automated Synthesis and Verification of Distributed Systems
Framework combining formal verification, program synthesis, and machine learning to automatically generate verified distributed system implementations. Achieves 274x speedup with 60% latency reduction through compositional verification and proof caching.
ETLC: A Context-First Approach to Data Processing in the Generative AI Era
A comprehensive whitepaper introducing ETLC (Extract, Transform, Load, Contextualize), adding semantic, relational, operational, environmental, and behavioral context to data pipelines.
Field-Theoretic Context System (FTCS)
An innovative approach modeling context as interacting fields rather than discrete states, enabling natural context flow and dynamic evolution through partial differential equations.
ARTEMIS - Adaptive Multi-agent Debate Framework
Technical disclosure on an adaptive framework for multi-agent decision systems using structured debate protocols to enhance enterprise decision-making.
Data Monetization Strategy for Enterprises
A comprehensive framework for enterprises to transform data into economic value, establishing methodologies now implemented across multiple JAPAC organizations.
OConsent: Open Consent Protocol for Privacy and Consent Management with Blockchain
A blockchain-based protocol for transparent personal data processing, enhancing user control and compliance with data privacy regulations.
Skills & Technologies
Technology Leadership & Strategy
Data Engineering & Architecture
Generative AI & Machine Learning
Cloud Platforms & Infrastructure
Programming & Development
Notable Projects
AI Metacognition Toolkit
Activation-level detection of sandbagging, deception, and situational awareness in LLMs. Linear probes achieve 90-96% accuracy across Mistral, Gemma, and Qwen models. Includes steering vectors for runtime behavior control.
- Sandbagging detection via linear probes with 90-96% accuracy
- Steering vectors reduce sandbagging behavior by 20%
- Bayesian situational awareness detection with KL divergence
- Available on PyPI • 275 tests • 95% code coverage
Open Location Proof Protocol
A privacy-aware open protocol for non-repudiable location verification in physical or virtual spaces.
- Cryptographically secure yet privacy-preserving protocol
- Fully decentralized architecture resistant to tampering
- Published comprehensive specifications for industry adoption
OConsent - Open Consent Protocol
An open-sourced transparent, fast, and scalable protocol for managing user consent and privacy on public blockchains.
- Blockchain-based solution for transparent consent management
- GDPR-compliant with automated audit capabilities
- Granular control over personal data usage
ISO8583 Simulator
A high-performance Java-based simulator for ISO 8583 financial messaging, used by banks and payment processors.
- Supports multiple ISO 8583 versions and custom formats
- Advanced performance testing with configurable load profiles
- Modular architecture supporting multiple protocols