SHEET 01 · WORK

Production GenAI at enterprise scale.

Enterprise AI platforms, practical applications and reusable engineering tools, built to connect architectural decisions with real-world outcomes.

SHEET 02 · FLAGSHIP WORK

Flagship work

100%of GenAI traffic

100+production use cases

FrameworkIn production

GenAI Service Layer

One managed surface for enterprise GenAI traffic.

When every team needs LLMs, you need infrastructure that turns chaos into reliability. A single managed surface for 100+ production use cases, multi-cloud, multi-vendor, model-agnostic across OpenAI, Anthropic, Google, AWS and more.

  • Shared infrastructure: model-agnostic routing, multi-cloud access and prompt/response caching support 100+ production use cases.
  • Governed operations: observability, audit trails, policy enforcement, quotas and chargeback give teams control over usage and cost.
  • Team enablement: client SDKs, guided administration with server-enforced approvals, spend analysis and a prompt/model workbench support platform adoption.

PLATE · 60s explainer · Captioned, music only

Video explainer: GenAI Service Layer.

16w → 2wor less, idea to production

FrameworkIn production

Agent Service Layer

A faster full idea-to-production pipeline, including evaluation, validation, load testing and deployment.

Agents need a repeatable path from idea to operation. Built atop the GenAI Service Layer, this framework standardises the agent lifecycle, from orchestration and memory to evaluation and deployment.

  • Reusable foundations: a standard agent harness, orchestration, parallel tool calls, memory and retrieval give teams shared building blocks.
  • Human oversight: approval checkpoints and regression evaluation support validation throughout the lifecycle.
  • Faster delivery: the full idea-to-production pipeline, including evaluation, validation, load testing and deployment, takes 2 weeks or less, compared with 16 weeks before.

PLATE · 63s explainer · Captioned, music only

Video explainer: Agent Service Layer.

85K+queries per week

80%+customer satisfaction

ProductIn production

Multi-agent customer chatbot

Customer satisfaction rose from ~50% to 80%+ over the program lifecycle.

The first proof that the Agent Service Layer scales: a customer-facing chatbot handling 85,000+ queries per week, with CSAT lifted from ~50% to 80%+ over the program lifecycle.

  • Coordinated agents: specialised agents delegate work and call tools in parallel to handle customer requests.
  • Connected journeys: persistent memory and real-time knowledge retrieval support booking, modification and support across multiple turns.
  • Human escalation: approval and escalation checkpoints keep people involved in high-stakes customer flows.

PLATE · 58s explainer · Captioned, music only

Video explainer: Multi-agent customer chatbot.

GoldAsian Design Awards 2024

Award

AI-powered travel-planning experience

GenAI-powered search doubled customer satisfaction on the search experience.

Travel planning combines GenAI search with smart flight recommendations, built on the GenAI Service Layer. The experience connects shared AI infrastructure to measurable improvements in customer satisfaction and engagement.

  • AI search: GenAI-powered search doubled customer satisfaction on the search experience.
  • Flight recommendations: the recommendation surface averages 2,000 queries a day, with click-through rising from 23.8% to 34.4% in three months.
  • Design recognition: the experience received Gold at the Asian Design Awards 2024.

PLATE · 49s explainer · Captioned, music only

Video explainer: AI-powered travel-planning experience.

SHEET 03 · HOW THE PLATFORM FITS TOGETHER

How the platform fits together

Enterprise GenAI platform architecture Layered platform. Top tier shows two use case classes: agentic use cases (which flow through the Agent Service Layer) and direct GenAI use cases (which call the GenAI Service Layer directly). The Agent Service Layer contains agent runtime, agent memory (session, short-term, long-term), agent auth for attended and unattended flows, tools and APIs and MCPs registry, multi-agent orchestration, planner and reasoning, human-in-the-loop, and trace and replay and eval. The GenAI Service Layer foundation contains smart routing, multi-cloud abstraction, caching, streaming and structured output, RAG and vector databases, and model registry. Cross-cutting platform capabilities wrap everything: observability, cost governance, access control, policy and compliance, infra management, audit, safety, secrets management, data residency, SLA and reliability, and disaster recovery. AGENTIC USE CASES DIRECT GENAI USE CASES Agentic use case Agentic use case … many more GenAI use case GenAI use case … many more Agent Service Layer Agentic systems framework Agent Runtime exec · lifecycle · sandbox Agent Memory session · short · long Agent Auth attended + unattended flows Tools · APIs · MCPs registry · invoke · parallel Orchestration multi-agent · handoffs Planner & reasoning ReAct · reflection · plans Human-in-the-loop approvals · checkpoints Trace · replay · eval agent-level test harness GenAI Service Layer Model-agnostic foundation · multi-cloud · multi-vendor Smart routing provider · model · tier Multi-cloud · multi-vendor OpenAI · Anthropic · Google · AWS Caching & token mgmt prompt · response · cost optim Streaming & schema tokens · structured output RAG & VectorDB embeddings · retrieval · hybrid Model registry versioning · catalog · canary CROSS-CUTTING PLATFORM CAPABILITIES Observability & tracing Cost governance Access control & IAM Policy & compliance Infra & capacity mgmt Rate limiting & quotas Audit & lineage Safety & guardrails Secrets & data privacy Data residency SLA & reliability Disaster recovery
Use cases Agentic + direct GenAI · 100+ in production
↓
Agent Service Layer Runtime · Memory (session, short, long) · Auth (attended + unattended) · Tools, APIs, MCPs · Orchestration · Planner · HITL · Trace & eval
↓
GenAI Service Layer Smart routing · Multi-cloud · Caching · Streaming & schema · RAG & VectorDB · Model registry
↓
Cross-cutting capabilities Observability · Cost · Access · Policy · Infra · Audit · Safety · Secrets · Residency · SLA · DR · Rate-limiting
Agentic flows go through the Agent Service Layer. Direct GenAI flows call the foundation directly. Cross-cutting capabilities apply to both.

SHEET 04 · BROADER PORTFOLIO

Broader portfolio

16 of 16 projects

SHEET 05 · PLATFORMS AND GOVERNANCE

Platforms and governance

Shared infrastructure and controls that help teams build and operate enterprise AI.

FrameworkProof of concept

Policy-Enforced Agent Authorization

A reusable authorization layer controls who can invoke an agent and what it may do on their behalf. Authority is bounded by the intersection of caller, agent and tool permissions.

  • Bounded authority: effective permissions are the intersection of caller, agent and tool authority.
  • Policy enforcement: delegated identity checks govern both agent invocation and tool use.
  • Progressive approval: identity verification and human approval/resume flows support actions that require stronger authorization.
FrameworkDeployed internally

Workload-Based Model Evaluation

An evaluation platform samples application workloads, replays them across candidate models and compares quality, cost and latency at task level. Model selection becomes a traceable decision grounded in the work an application actually does.

  • Representative replay: sample application workloads and compare candidate models on the tasks they will actually perform.
  • Decision evidence: task-level quality, cost and latency comparisons make model tradeoffs explicit.
  • Human calibration: traceable judgments and judge-agreement checks focus review on decisions that need it.
FrameworkDeployed internally

Unified Enterprise AI Workspace

A shared workspace brings AI applications behind one sign-in and centrally managed access. Separating core identity from optional organizational context keeps permissions reliable when supporting integrations fail.

  • Shared access: single sign-on and role-based permissions give users a common entry point to multiple applications.
  • Resilient identity: access decisions depend on authoritative identity, independently of optional directory enrichment.
  • Accountable publishing: human review precedes publication, with readback checks to verify the published result.

SHEET 06 · APPLIED AI SYSTEMS

Applied AI systems

Applying AI to customer journeys, operational workflows and everyday decisions.

ProductDeployed internally

Evidence-Gated Document Extraction

Document extraction needs evidence behind each field. This engine links extracted values to source material and applies deterministic verification to guide release, recovery and human review.

  • Difficult layouts: adaptive page subdivision and overlap stitching help recover tables and rows.
  • Targeted recovery: verification failures guide follow-up extraction toward missing or unsupported values.
  • Evidence-led review: abstention and human review handle fields where the available evidence is insufficient.
ProductDeployed internally

Reusable Spreadsheet Workflows

A spreadsheet agent turns conversational requests into bounded server-side data operations. Completed flows become parameterized workflows that can be reused with new inputs.

  • Conversational analysis: structured tool calls translate requests into server-side spreadsheet operations.
  • Repeatable work: save completed sequences as parameterized workflows and replay them with new inputs.
  • Controlled extensions: role checks and separately approved sandbox execution support tasks beyond the standard operation set.
ProductDeployed internally

Governed AI Classification

Business categories change, and classification systems need to make those changes traceable. A workbench turns category definitions into reviewed, versioned decision rules with evidence behind the outputs.

  • Versioned rules: immutable releases and validation gates make changes to classification criteria explicit.
  • Recoverable processing: durable batch jobs support classification work that can resume after interruption.
  • Traceable decisions: evidence-linked outputs retain human review of classification results.
ProductPrototype

AI-Assisted Legacy Modernization

Modernizing an operational system starts with preserving its system of record. Versioned changes, reconciliation and record-grounded AI insights introduce new capabilities while keeping their relationship to existing data explicit.

  • Controlled change: versioned updates and reconciliation connect new workflows to existing records.
  • Grounded statements: models select constrained references and categories; application code constructs displayed statements from underlying records.
  • Reviewed insights: human review connects AI-assisted operational insights to the records that support them.
ProductDeployed demonstrator

Structured Travel Query Extraction

A travel-query extractor turns free-text requests into structured flight-search parameters. Typed model answers capture intent; deterministic code resolves destinations, passenger counts and concrete dates.

  • Structured intent: a fixed question schema and guardrail checks turn free-text requests into typed search parameters.
  • Precise resolution: focused follow-up addresses ambiguous destinations; application code resolves passenger counts and dates.
  • Visible performance: per-extraction latency, token usage and cost make the processing tradeoffs measurable.
ProductPackaged application

Synthetic Media Provenance

Assessing synthetic media requires evidence about its origin. This application examines images and document imagery through signed provenance and watermark signals, with reports that distinguish detection, absent signals and uncertain checks.

  • Provenance evidence: signed content credentials and watermark checks retain the basis for each finding.
  • Document coverage: examines both standalone images and imagery embedded in documents.
  • Clear interpretation: separate reporting of detected generation, absent signals and uncertain checks supports informed review.

SHEET 07 · KNOWLEDGE AND RESEARCH

Knowledge and research

Turning documents and scattered information into traceable answers, maintained knowledge and actionable research.

Research & toolingHosted app deployed

DocIQ: Document Intelligence

Answering questions about complex documents requires finding and inspecting the right pages. DocIQ combines lexical search, agent-directed exploration and visual inspection in a reusable retrieval engine.

  • Page-grounded answers: lexical search locates source pages, while visual inspection captures evidence in tables, diagrams and scans.
  • Adaptive retrieval: agents expand searches and inspect additional pages as needed, with configurable provenance and citations.
  • Reusable integrations: server, agent-tool and hosted application forms adapt the same retrieval approach to different workflows.
Research & toolingLocal application

Tacit: Curated Organizational Knowledge

Scattered source material becomes useful organizational knowledge when people can trace, review and maintain it. Tacit connects source-linked knowledge notes to curated topic articles and human ownership.

  • Traceable curation: source-linked knowledge notes feed maintained topic articles.
  • Conflict detection: competing or outdated information is surfaced for review as knowledge changes.
  • Human ownership: topic owners and a review queue guide what becomes maintained organizational knowledge.
Research & toolingCloud edition deployed

SmartResearch: Steerable Research

The inverse of Google's Deep Research: research with room for course correction before the final report. An eight-stage workflow uses editable, versioned checkpoints so users can inspect, revise and approve the work as it develops.

  • Human checkpoints: review the question, plan, sources, findings and report throughout the eight-stage workflow.
  • Focused revision: selective section regeneration and document export turn reviewed findings into a usable report.
  • Research workspaces: the cloud edition supports web research with isolated user workspaces and provider choice; the original edition pairs web research with private-document retrieval.

SHEET 08 · ENGINEERING ENABLEMENT

Engineering enablement

Reusable workflows and portable knowledge for AI-assisted development.

Research & toolingPackaged tooling

Reusable Agent Skills and Portable Knowledge

AI-assisted development needs consistent workflows and knowledge that travels between projects. Reusable skills capture how to work, while Open Knowledge Format (OKF) bundles keep source-linked knowledge available to coding agents.

  • Shared workflows: project setup, session handoff, release review and API verification capture repeatable engineering practices.
  • Portable knowledge: plain-text bundles record sources, trust and freshness, with indexes that let agents retrieve relevant context as needed.
  • Reliable distribution: reproducible packages, checksums, installation safeguards and offline tests support reuse across projects.

SHEET 09 · PARTNERSHIPS AND ADOPTION

Partnerships and adoption

Industry collaboration, organizational capability and emerging interfaces.

PartnershipsCross-industry collaboration

Industry partnerships

Frontier AI capabilities become enterprise value through close technical collaboration. Engagement with leading AI organizations connects pre-release evaluation and roadmap input to customer service, multimodal experiences and staff assistance.

  • OpenAI: Industry-first major-airline collaboration covering enterprise chatbot, multimodal customer servicing, and staff AI assistance.
  • Anthropic: Enterprise readiness collaboration around Claude in production, agentic systems, and safety-first deployment patterns.
  • Google: Frontier model access, multimodal capabilities, and platform-level integration via Vertex AI.
  • AWS: Bedrock-based multi-vendor model access, enterprise compliance, and infrastructure scaling.
EngagementCross-divisional · C-suite reach

Building Enterprise AI Capability

Enterprise AI adoption depends on people who can build, use and lead it. I design and lead programs that connect hands-on engineering practice with product decisions and executive understanding.

  • Practical learning: hands-on workshops and upskilling help engineering and product teams adopt GenAI.
  • Technical mentorship: individual coaching develops agentic design skills in senior engineers, architects and emerging technical leaders.
  • Organizational reach: executive briefings and train-the-trainer programs extend AI capability across business divisions.
Voice / MultimodalNative streaming voice

Voice AI for Staff Operations

Staff workflows need conversation that responds in real time. Native streaming voice brings low-latency, prosody-aware interaction to operational tasks.

  • Native voice: end-to-end speech interaction replaces separate speech-to-text and text-to-speech stages.
  • Real-time response: streaming supports low-latency conversation during operational workflows.
  • Expressive interaction: prosody-aware conversation carries vocal cues through the staff experience.
Research directionsOther active threads

Further research

Further research explores how AI systems can become more resilient, precise and useful:

  • Adversarial robustness and guardrails: Defence-in-depth against prompt injection, jailbreaks, data exfiltration, and model-targeted attacks. Continuous red-teaming, not one-off audits.
  • Knowledge graphs + LLM hybrid systems: Structured retrieval and reasoning over enterprise knowledge, complementing or replacing pure vector RAG where precision matters.
  • On-device and edge inference: Local-first GenAI for sensitive, regulated, or low-connectivity workflows where data cannot leave the device.

For a specific topic or collaboration not listed here, the fastest way to start a conversation is LinkedIn.