Jeriel Chan Zhi Yang

Markov chains, martingales, Poisson processes, and Brownian motion applied to financial mathematics, queueing models, and statistical inference.

Theoretical foundations of RL: Markov Decision Processes (MDPs), Bellman optimality operators, policy/value iteration, policy gradients, Q-learning convergence proofs, and PAC-MDP sample complexity bounds.

Quantum state spaces, qubits, quantum logic gates, quantum circuits, Grover's search algorithm, Shor's factoring algorithm, and quantum error correction fundamentals.

Distributed consensus protocols (Paxos, Raft), consistency models (linearizability, eventual consistency), mutual exclusion, distributed transactions, and parallel algorithm design.

Partially Observable Markov Decision Processes (POMDPs), classical STRIPS/PDDL planning, heuristic search (A*, Monte Carlo Tree Search), and multi-agent temporal planning.

NUS Mountaineering — Expedition team; summited Tsorku Peak (5,800 m), Himalaya, Nepal.

NUS Calisthenics — Competitive streetlifting athlete.

NUS Computing Code For Community — President; led 30 mentors providing free coding lessons to underserved children.

Athena is JPMorgan's proprietary Python-based front-office trading platform, handling real-time pricing, risk, and trade lifecycle management for Fixed Income, FX, and Equities — one of the largest in-house trading systems in finance.

The Emerging Markets (EM) fixed income desk focuses on bonds and rates in developing economies, requiring very low-latency pricing, hedging, and inventory management.

The Model Context Protocol (MCP) is an open standard by Anthropic letting LLMs call external tools and data sources in a structured, safe way.

Key design decisions:

• Role-based entitlements: Users (trader, risk, ops) could only invoke tools they were authorized for — enforced at the MCP server layer, not just the UI.

• Guardrails: Output filters prevented hallucinated numbers or unauthorized assertions — critical in a regulated financial environment.

• SPA bundling: Packaged as a single-page app for embedding into existing internal portals.

Received line-of-business approval — passing compliance, legal, and technical review for production use.

Migrated legacy internal dashboards to modern TypeScript + React frontend backed by Java Spring Boot microservices. Involved complex financial table layouts with real-time data and ensuring backward API compatibility.

Phoenix is the Home Team's first internal LLM chatbot, deployed across 11 sub-agencies (Police, Civil Defence, Immigration, Prisons, etc.).

• Developed with Mistral AI and Microsoft, deployed on NVIDIA H100 GPUs.

• RLHF fine-tuning: Curated domain datasets via web scraping + preprocessing, then preference-based fine-tuning aligned with government communication norms.

• Safety guardrails preventing inappropriate content in a sensitive law enforcement context.

• Presented outcomes to Ministers and Directors.

Full-stack multimodal poster tool for the Central Narcotics Bureau (CNB):

• Next.js frontend with drag-and-drop template selection and live preview.

• Azure OpenAI (GPT-4o) for copywriting from campaign briefs.

• LlamaIndex for RAG over past designs and brand guidelines.

• Flux (image diffusion) for generating visual assets.

• MongoDB for template storage and sessions.

Result: Reduced poster creation from hours to minutes.

Augmodo builds AR shelf-scanning tools for retail. CV model accuracy metrics (object detection, planogram compliance) tracked per SKU and store location.

• Vue 3 with reactive state and custom charts.

• Firestore caching with TTL invalidation — reduced read costs; write buffer batched metric updates from edge devices.

• Surfaced false-positive rates and confidence distributions enabling PMs to identify model drift.

Led 11 remote contractors across time zones for CV/AR annotation tasks.

• Wrote labeling guidelines and edge-case docs for inter-annotator consistency.

• Computed inter-annotator agreement (Cohen's kappa) and flagged low-confidence samples.

• Weekly syncs and task tracking to hit labeling quotas on time.

QMAVIS = Query-Mapped Audio-Visual Instruction Synthesis. A fusion architecture routing queries to specialized sub-models (audio, visual scene, temporal reasoning) and aggregating via learned weighting — enabling coherent understanding of very long videos that exceed single-model context windows.

VideoMME: Industry-standard multi-modal eval on short/medium/long videos (up to 1hr+), with and without subtitles.

PerceptionTest: DeepMind benchmark for visual perception — object tracking, depth ordering, counting over short clips.

EgoSchema: Meta's first-person video QA — 5,000 MCQs over 3-minute clips requiring long-term temporal reasoning.

VideoLLaMA2: Strong open-source baseline from Alibaba DAMO Academy that QMAVIS outperforms.

Data structures: std::map<Price, Level> sorted by price; each Level is std::deque<Order*> — O(log n) price lookup, O(1) FIFO dequeue for time-priority matching.

O(1) cancel: Orders store index into their deque. unordered_map<OrderID, Order*> gives O(1) lookup. Cancel is direct deque erase.

Fixed-point arithmetic: Prices as integers (x100) to avoid floating-point rounding causing erroneous crosses.

Order types: ELO (Enhanced Limit), SLO (Stop-Limit, triggered at stop price then rests as limit), AON (All-or-None, full match or reject), Market (walks book at best price).

The MOS 6502 is an 8-bit CPU from 1975 powering the NES, Apple II, Commodore 64. Emulating it accurately requires all 56+ official opcodes plus undocumented opcodes real games relied on.

• Memory-mapped bus: ROM, RAM, PPU registers, APU, and controllers share a 16-bit address space — reads/writes dispatch by address range.

• Cartridge mapper support: Games used hardware bank-switching chips (iNES mapper numbers) to address more ROM than the address space allowed.

• Cycle accuracy: Instructions take varying cycles — critical for PPU synchronization (the graphics chip is cycle-coupled to the CPU).

Traditional ramps require trusting a centralized exchange. FlareGate replaces this:

1. Buyer locks crypto into smart contract escrow on Flare Network.

2. Seller sends fiat via any payment rail (PayNow, bank transfer, etc.).

3. Flare's Data Connector (on-chain oracle) cryptographically attests the fiat payment — no trusted intermediary.

4. Smart contract automatically releases escrowed crypto upon verified attestation.

Result: Trustless settlement with no centralized exchange risk.

Omega City gamified coding education by wrapping programming challenges inside an MMORPG game world. Players completed “quests” that were actually coding exercises.

• NUS VIP: Competitive grant requiring full business plan, market analysis, and VC pitch panel.

• HANGAR: NUS startup incubator with office space, mentorship, and investor introductions.

• ACRA-registered as a formal Singapore company.

SMO: National competition by Singapore Mathematical Society. Tests combinatorics, number theory, geometry beyond curriculum. Bronze = top ~15–20% nationally.

AMC: International — 100,000+ participants in 30+ countries. Distinction = top ~15% globally.

Euclid Contest (CEMC, Waterloo): Pre-university proof-writing and advanced problem-solving. Distinction = top ~25% internationally.

The Special Operations Command (SOC) is a specialized SPF formation for public order maintenance, riot control, VIP protection, and emergency response. Officers trained in crowd management, use-of-force protocols, and tactical operations under pressure.