Philip Cheung

Principal Forward Deployed Engineer

Applied AI · LLM Systems in Production · Financial Services

MSc Computer Science, AI & Data Science (Merit)

London, United Kingdomgithub.com/pakmingcwww.Philip.pmLinkedInBook a 30-min chat

Professional Summary

I am the engineer who gets AI from a working demo into a business that has to live with it. As Principal Forward Deployed Engineer at a Swiss corporate finance advisory firm I run a five-person team and personally own the architecture, integrations and production accountability across four products, reporting weekly to a finance-led board. I build with AI coding agents under a three-gate review and evaluation process, which is how a small team ships integrations with Open Banking, Stripe, Firebase and twelve LLM providers, and I carry the pager when it breaks. Current role: Principal Forward Deployed Engineer, Weidmann & Cie. AG, London. MSc Computer Science, AI and Data Science (Merit).

Core Capabilities

Discovery to Production Ownership

SLAs, audit trail and rollback path defined before go-live
Escalation point and on-call for production incidents
Scoped, dated deliveries, not demos
Weekly reporting to a finance-led board

Integration Depth

UK Open Banking, read-only and PII-minimised
Stripe: subscriptions, credit packs, usage-based billing
Firebase / Firestore, Cloudflare (Pages, Functions, Workers, D1, R2), AWS
FIX connectivity for the trading desk
12 LLM providers, Groq / Cerebras LPU inference

Governance Built In

Policy gates and evidence packs
Prompt registry and model routing
Per-call cost ledger and audit ledger
Human-review escalation for high-risk cases

Adoption and Stakeholder Work

Thin working prototype plus live dashboard instead of slide decks
Board, investor and regulatory requirements into dated deliveries
Weekly board reporting
Client-side background: enterprise account management and technical product management

Deployment Leadership

Five-person team across development and financial analysis
8 engineers mentored at Pacific Cloud
Seniors ship v1, juniors iterate under review
Architecture reviews and AI/ML working practices

How I Deliver (AI-Assisted)

Claude Code, Codex CLI, Gemini CLI, Cursor
Eval harness and three review gates
Spec first; agents implement; I review, debug and own the result
12-provider benchmark grounds model, latency and cost choices

Professional Experience

Principal Forward Deployed Engineer

Weidmann & Cie. AG

London, UK (Hybrid)December 2024 - Present

Swiss corporate finance advisory firm building an applied-AI product portfolio: Qorinix (ultra-low-latency AI inference cloud), LaSpend (read-only AI money assistant over UK Open Banking), Fixxmi (Swiss consumer service marketplace) and an AI-native trading desk. I take each product from discovery to production with a five-person team and report weekly to a finance-led board.

Own end-to-end architecture and production accountability across all four products on AWS, Cloudflare (Pages, Functions, Workers, D1, R2) and Firebase; I am the escalation point for production incidents
Designed and shipped the Qorinix inference control plane with AI coding agents under a three-gate review and evaluation process: multi-tier model router across 12 providers (OpenAI, Anthropic, Gemini, DeepSeek, Qwen, Groq, Cerebras and others), streaming SSE, per-call cost ledger, usage-based billing and entitlements; targets sub-200ms time-to-first-token
Integrated LaSpend with regulated UK Open Banking data, Stripe and Cloudflare D1: subscription detection, waste scoring and assisted cancellation, designed read-only by policy with human approval for every user-triggered action and deterministic fallback when models fail
Delivered Fixxmi's AI lead matching (Next.js 15, Firebase Cloud Functions, Firestore europe-west6) across 12+ service categories in DE-CH/EN within GDPR and nFADP boundaries, with pay-per-lead Stripe credit packs
Built the governance loop applied to every AI call: policy check, evidence-pack retrieval, prompt registry, model routing, audit ledger and human-review escalation for high-risk cases
Ran a 12-provider inference benchmark (LLM Arena, live at philip.pm) to ground model, latency and cost decisions; introduced token tiering and self-hosted models for non-latency-critical workloads to cut API spend
Translate board, investor and regulatory requirements into scoped, dated deliveries; win decisions with thin working prototypes and live dashboards rather than slide decks

AI Engineer & Technical Architect

Pacific Cloud Computing Ltd.

Hong Kong & Remote UKDecember 2021 - December 2024

Promoted internally from Senior Software Engineer to lead the firm's AI transformation across SaaS, analytics and market-intelligence products.

Built and operated a real-time ML inference service (feature pipelines, model serving, monitoring) for the firm's analytics products, with latency and accuracy gates defined before each release
Implemented enterprise retrieval (RAG) and market-intelligence NLP pipelines, including the evaluation methodology used to accept them into production
Directed the launch of three multi-tenant SaaS platforms; ran architecture reviews and established AI/ML working practices
Mentored an agile team of 8 engineers: seniors ship v1, juniors iterate under review
Led quantitative-markets decision support research: cross-asset anomaly detection and price forecasting

Senior Software Engineer

Pacific Cloud Computing Ltd.

Hong KongJanuary 2015 - November 2021

Built an enterprise document management SaaS with real-time collaboration for corporate clients
Built e-commerce platform capabilities including multi-currency payments; the SEO programme reached top search rankings
Pioneered the firm's blockchain product work: smart-contract products, NFT marketplace support and DeFi analytics

Senior Product Manager (Technical) / Web & Content Manager

Groupon.com

Hong KongApril 2013 - December 2014

Led platform modernisation from a monolithic to a service-oriented architecture as technical product manager, coordinating engineering delivery and business stakeholders
Introduced an A/B testing and SEO programme that delivered 150% traffic growth
Ran digital product operations and commercial prioritisation for the marketplace: merchandising, pricing structure and merchant coordination in an Agile environment

Senior Operations Manager

SoManyCall Telecom

Hong KongMarch 2008 - April 2013

Oversaw strategic and operational delivery for custom software solutions in the telecoms sector, achieving a 25% annual growth rate
Led a multi-disciplinary team across the full project lifecycle

Core Technical Competencies

AI-Assisted Engineering

Claude Code

Codex CLI

Gemini CLI

Cursor

Spec-Driven Delivery

Eval Harness

Three Review Gates

Agent Output Review & Debugging

Responsibility Tags

Generative AI & LLM Systems

Multi-Provider Routing

RAG

Eval Harness

Streaming SSE

Prompt Registry

Guardrails

Function Calling

Token Tiering

Self-Hosted Models

Integrations & Payments

UK Open Banking

Stripe

Firebase / Firestore

Usage-Based Billing

Credit Packs

Webhooks & Idempotency

FIX

Groq / Cerebras LPU

Cloud, DevOps & Infrastructure

Cloudflare

AWS

GCP

Docker

Pages / Functions / Workers

D1 / R2

Kubernetes

Terraform

CI/CD

OpenTelemetry

Multi-Region Failover

Backend, APIs & Microservices

TypeScript

Next.js

FastAPI

Node.js

React

REST APIs

Microservices

Event-Driven

Edge Runtime

Data Pipelines & Storage

PostgreSQL

Redis

TimescaleDB

pgvector

Firestore

Cloudflare D1 / R2

Batch Pipelines

Real-Time Pipelines

Machine Learning & Deep Learning

Python

PyTorch

Scikit-learn

XGBoost

TensorFlow

Transformers

Decision Trees

LSTM/GRU

Fine-tuning

Feature Engineering

MLOps & Production ML

Model Evaluation

A/B Testing

Model Serving

Automated Retraining

Canary Deployment

Drift Monitoring

ONNX

Triton

Quantisation

Selected Deployments

Qorinix Inference Control Plane

AI agents under my review

Problem: The board wanted "instant" AI responses for real-time agents, voice and trading alerts, inside a latency budget, a provider cost ceiling and with failover when any single provider degrades.
What shipped: A multi-tier model router across 12 providers (OpenAI, Anthropic, Gemini, DeepSeek, Qwen, Groq, Cerebras and others) with streaming SSE, a per-call cost ledger, usage-based billing and entitlements; token tiering and self-hosted models for non-latency-critical work.
Outcome and measurement: Targets sub-200ms time-to-first-token (p50) and 100-200+ tokens/sec, measured by TTFT p50/p95 per provider in the LLM Arena benchmark. Token tiering cut API spend for non-latency-critical work (no percentage published).
My role: Designed and shipped with AI agents under my review

LaSpend Open Banking

AI agents under my review

Problem: A read-only AI money assistant over FCA-regulated UK Open Banking data: subscription detection, waste scoring and assisted cancellation, with PII minimisation and a rule that it never moves money.
What shipped: Integrations with the Open Banking API, Stripe and Cloudflare Pages, Functions and D1; human approval for every user-triggered action, deterministic fallback when models fail and verified action receipts.
Outcome and measurement: Live in production. Measured by detection precision on recurring spend and cancellation completion rate (user numbers not published).
My role: Designed and shipped with AI agents under my review

Fixxmi AI Lead Matching

AI agents under my review

Problem: A Swiss consumer service marketplace needed AI lead matching across 12+ categories in DE-CH/EN within GDPR and nFADP, with Firestore europe-west6 data residency.
What shipped: Next.js 15, Firebase Cloud Functions and Firestore (europe-west6) with intent classification and matching inside the governance loop; pay-per-lead monetisation with Stripe credit packs.
Outcome and measurement: Live. Measured by match acceptance rate (volume numbers not published).
My role: Designed and shipped with AI agents under my review

Governance Loop

Built

Problem: Four products with different risk profiles needed one control layer that a finance-led board and a regulator could both read, with every decision reconstructable after the fact.
What shipped: A loop applied to every AI call: policy check, evidence-pack retrieval, prompt registry, model routing, action runtime, audit ledger, human-review escalation and policy feedback.
Outcome and measurement: Measured by audit completeness and escalation rate; controls tighten from observed production behaviour.
My role: Built

Portfolio Assistant (philip.pm)

Built

Problem: A public portfolio chatbot that answers hiring questions instantly, cannot be hijacked and costs almost nothing to run: the live, inspectable proof of hands-on work.
What shipped: Next.js edge runtime on Cloudflare Pages, a five-provider cascade with timeouts and fallback, signed-cookie rate limiting with staged daily budgets, a curated preset answer bank streamed instantly and a hardened system prompt.
Outcome and measurement: Live at www.philip.pm. Measured by fallback rate and p95 response time.
My role: Built

Education

MSc Computer Science, AI & Data Science

University of Wolverhampton, UK

2023 - 2025 | Grade: Merit

Dissertation: LLM-Augmented High-Frequency Trading Strategy Development
AI-assisted three-gate validation workflow: 85% reduction in strategy development time
Modules: Deep Machine Learning, Intelligent Agents, Data Science & Mining, Applying AI, Cloud Computing, Research Methods

Bachelor of Business Administration

Hong Kong University of Science & Technology

1995

Marketing with Information Systems minor

Selected Certifications

MLOps SpecializationGoogle Cloud | 2024

Gemini Certified EducatorGoogle | 2025-2028

CS50x Computer ScienceHarvard | 2024

SFC Licensing Exams (Papers 1/7/8/12)HKSI, Hong Kong | 2021

Open to Principal Forward Deployed, Applied AI and AI Solutions Engineering roles

Full UK right to work | London, hybrid or remote UK | Languages: English, Cantonese, Mandarin

Book a 30-min chat

Last Updated: September 2026 | Portfolio: www.Philip.pm | References available upon request