
PHD APPLICATION PRESENTATION / V0.6
I build research systems that turn complex questions into testable evidence.
Pengyi Peng / complex systems, AI agents, quantitative research, and research engineering.

01 / RESEARCH IDENTITY
My research advantage is integration across models, systems, and evidence.
I move from mathematical abstraction to reproducible software, measurable experiments, and Human-reviewed decisions.
- Complex systems / model interactions and emergent behavior
- AI agents / memory, tools, workflow, evaluation
- Quantitative systems / markets, risk, data, and execution

02 / EDUCATION
Mathematics and complex systems trained me to model interactions, not isolated variables.
The academic foundation combines applied mathematics with a systems view of dynamics, networks, uncertainty, and computation.
- MSc Complex Systems Modelling / King's College London / 2022-2024
- BSc Mathematics and Applied Mathematics / University of Reading and NUIST / 2018-2022
- First-Class Honours / top 10% of programme / pending final CV fact review

03 / METHOD
I treat research as a governed loop from question to falsifiable result.
Problem -> system model -> hypothesis -> experiment -> benchmark -> review -> new question.
- Make state, assumptions, and invariants explicit
- Separate exploratory code from reference implementation
- Preserve failures and reviewer feedback as learning data

04 / QUANTITATIVE EVIDENCE
Global competitions tested my ability to convert ideas into ranked quantitative performance.
WorldQuant results provide external evidence of research iteration under common rules and competitive evaluation.
- International Quant Championship 2024 / global 232 of 34,142 / top 0.7% / UK 5
- MAPC 2024 / global 11 of 850 / UK 1
- Research habit / generate, test, compare, diagnose, iterate

05 / PROFESSIONAL EXPERIENCE
Banking and securities work grounded my systems thinking in real institutions and constraints.
The experience exposed how data quality, process ownership, disclosure boundaries, and cross-functional coordination shape financial decisions.
- Ping An Bank / management trainee / corporate, retail, and operations contexts
- CITIC Securities / quantitative private-fund comparison and manager communication
- Research implication / institutional workflows are socio-technical systems

06 / AI AGENT SYSTEMS
My Agent work focuses on reliable workflows, not impressive single prompts.
State, memory, tools, evaluation, provenance, and Human approval must operate as one research system.
- Versioned schemas and deterministic state transitions
- Append-only traces, regression tests, and explicit failure semantics
- Private evidence separated from public-safe artifacts

07 / OPEN-SOURCE RESEARCH ENGINEERING
Upstream review taught me to make small claims survive large codebases.
Merged fixes across AI retrieval, trading infrastructure, and knowledge systems required bounded changes, regression evidence, and maintainer dialogue.
- LightRAG / token-limit and query-boundary failure handling
- NautilusTrader / Rust test runtime without changing behavior
- WeKnora / archived-page semantics in active statistics

08 / QUANTITATIVE RESEARCH
I am converting scattered market ideas into reproducible research pipelines.
Data -> feature -> signal -> backtest -> evaluation -> risk -> implementation boundary.
- Futures / ATR, RSI, and term-structure signals
- Microstructure / L1 features and simplified market-making models
- Evaluation / assumptions, tests, regimes, failure analysis

09 / FIXED INCOME
Fixed income gives me a domain where mathematics, data, risk, and Agent research meet.
The current Rates Bond Quant laboratory connects duration, convexity, DV01, curves, and interactive explanation.
- Rates / term structure, carry, roll-down, relative value
- Risk / duration, convexity, DV01, scenario response
- Next research / factor robustness and Agent-assisted evidence synthesis

10 / RESEARCH INFRASTRUCTURE
I build the interfaces and infrastructure needed to make research inspectable.
Research software matters when another person can run it, question it, and trace a conclusion back to evidence.
- PWEB / React, TypeScript, Worker API, D1 persistence
- PINF / software and hardware constraints as one measured system
- Deployment / tests, CI, public-safe surfaces, versioned releases

11 / PROJECT PORTFOLIO
The portfolio is becoming a federated laboratory rather than a list of unrelated repositories.
Shared evidence, Agent workflows, benchmarks, and deployment patterns let each project improve the next one.
- Research / agent systems, quant, FICC, papers, experiments
- Engineering / polyglot coding, infrastructure, full stack, open source
- Delivery / CV, presentation, interview, communication, feedback

12 / CODING + SYSTEM DESIGN
Across languages, I look for the same system truths.
Data -> type -> state -> invariant -> transition -> interface -> boundary -> architecture.
- Python / fast research and Agent workflows
- TypeScript / full-stack contracts and product interfaces
- Rust, Go, C++, OCaml / systems, correctness, performance, and comparative design

13 / WORKING WITH A RESEARCH GROUP
I can contribute as both a researcher and the engineer who makes research usable.
My strongest contribution is translating uncertain questions into experiments, systems, artifacts, and clear review surfaces.
- Methods / quantitative modelling, backtesting, RAG and Agent evaluation
- Engineering / Git, Linux, APIs, testing, CI, Cloudflare, technical writing
- Communication / Mandarin, English, Cantonese; cross-functional and open-source review

14 / PHD RESEARCH AGENDA
I want to study how Agent systems can produce reliable evidence in complex, changing domains.
The initial agenda joins Agent architecture, evaluation, domain knowledge, and real-world decision constraints.
- Reliability / when do memory, retrieval, or tools improve decisions rather than add noise?
- Evaluation / how do we benchmark long-horizon, multi-step research behavior?
- Domain proving ground / finance, markets, and scientific research workflows

15 / FIRST 90 DAYS
I want the first collaboration to end with a benchmark, a reproducible artifact, and a sharper question.
Map the literature -> reproduce a baseline -> define failure cases -> run one bounded study -> release evidence for review.
- Choose one supervisor-aligned question
- Fix the dataset, baseline, protocol, and evaluation contract
- Publish code, report limitations, and decide the next experiment