Secret Route
← Return
STAND USER · スタンド使い
ARCANA: THE MAGICIANStand User

KEN WU

吴锵皓
SOFTWARE ENGINEERML RESEARCHERCS '26
LOCATION
Toronto, ON
BASE
University of Waterloo
SEEKING
Fall 2026 / New Grad
Stand Name
DEEP PURPLE
Stand Stats
DESTRUCTIVE POWER
B
SPEED
A
RANGE
A
DURABILITY
B
PRECISION
A
DEV POTENTIAL
∞
Battle History · 戦歴
Chapter 01 · AGENTIC LAGRANGIAN EXTRACTION

Fermilab

Google Summer of Code · ML4SCI · University of Alabama
May 2026 — Present
Remote · United States
Open Source Developer
▶Built agentic Lagrangian extraction tooling for ML4SCI at Fermilab under Google Summer of Code 2026 with University of Alabama mentors
▶Open-sourced the pipeline for physics-informed symbolic extraction; project details on the GSoC 2026 page
Python
Chapter 02 · ML · ATTENTION & SYMBOLIC REGRESSION

University of Waterloo

With Prof. Ali Ghodsi & Amin Ravanbakhsh
Sep 2024 — Apr 2026
Waterloo, Canada · Remote
Undergraduate Research Assistant
▶Held R² ≥ 0.99 on numerically stable symbolic-regression fits by benchmarking models at dataset scale
▶Raised in-domain accuracy ~19pp by fine-tuning Symbolic GPT variants under Prof. Ali Ghodsi
▶Balanced R² vs. model complexity via tokenizer and Point-Net ablations across benchmark suites
▶Cut MSE and MRE on standard benchmarks by refining inference loops for Symbolic GPT
PythonPyTorchTransformers
Chapter 03 · UNSUPERVISED LEARNING

Lancaster University

With Prof. Plamen Angelov
Jan 2025 — May 2026
Lancaster, UK · On-site
Undergraduate Researcher
▶Sped up PAMSil up to 85.6% on CIFAR-100 at equal quality with recursive ReSil / ReSilC (O(1) key updates)
▶Beat K-Means wall-clock by 17–24% on CIFAR-10/100, MNIST, and Fashion-MNIST via optimized R-Means centroids
▶Benchmarked recursive vs. flat clustering across 8+ datasets with 10-run averages in NumPy / scikit-learn
▶Kept speed-quality comparisons fair by tracking silhouette, inertia, and wall-clock on every run
PythonNumPyscikit-learn
Chapter 04 · ENTERPRISE AI

Mercor

Incoming
San Francisco, United States
Member of Technical Staff
▶Incoming Member of Technical Staff, Enterprise AI.
Chapter 05 · BAILOUT MODEL

Lyft

May 2026 — Aug 2026
Toronto, Canada
Machine Learning Software Engineer Intern
▶Doubled the bailout model’s cancel-detection recall (+34.2pp at 2.5% FPR) via 17 new route/GPS features
▶Grew Canada bailout coverage from 1.4% to 100% of rides via census backfill for 2,964 geohashes
▶Root-caused a defect in Lyft’s bailout model flagging 76.6% of rides; shipped automated per-cluster thresholds
▶Shipped shadow inference and holdout capture into the bailout service with zero live decision impact
PythonCatBoostSQLGeoPandas
Chapter 06 · 5G NETWORK AGENT

Nokia

Jul 2025 — Dec 2025
Ottawa, Canada
Software Engineer Intern
▶Shipped a LoRA fine-tuning pipeline that turns Qwen into a Camunda BPMN XML generator with GGUF quantization
▶Caught NCSC-only training overfitting with an eval harness that benchmarks fine-tuned vs. base models
▶Unblocked GPU training by fixing OOM and Triton kernel failures; shipped reusable training tooling
PythonPyTorchUnslothTransformer
Chapter 07 · INSURANCE ANALYSIS

TD Bank

Jun 2025 — Aug 2025
Toronto, Canada
Data Scientist Intern
▶Replatformed 1.5M+ row pipelines and reproduced a 154,340-row deliverable at 100% parity
▶Cut QA time 80% by automating 30+ min/cycle ingestion with a row-level parity harness
▶Saved 2–4 hrs/week by migrating the on-prem pipeline to Azure with automated runs
PythonSQLPySparkPandasDatabricks
Chapter 08 · TEACHING

Stanford University

Apr 2025 — Jun 2025
Stanford, United States
Student Instructor
▶Instructed Stanford Code in Place CS106A for a global cohort of 40,000+ students
▶Taught intro Python with Karel and Tkinter through live sections and project feedback
PythonKarelTkinter
Chapter 09 · LLM AGENT

August

Sep 2024 — Dec 2024
New York, United States
Software Engineer Intern
▶Scaled 10+ FastAPI / AWS / Supabase endpoints to 2,000+ requests/min in production
▶Cut eval cycles across 15+ LLM agents with round-robin multi-agent routing and ELO scoring
▶Architected end-to-end multi-agent RAG with LLM-as-Judge evaluation
PythonLangGraphFastAPIAWS
Chapter 10 · SUPER RESOLUTION

hum.ai

Formerly Coastal Carbon
Jul 2024 — Sep 2024
Kitchener, Canada
Machine Learning Engineer Intern
▶Benchmarked SOTA super-resolution models (ESRGAN, StableSR) in PyTorch for coastal imagery
▶Automated multi-model SR eval pipelines on SageMaker with Matplotlib / Seaborn analysis notebooks
▶Stood up AWS S3 / EC2 experiment infra for scalable fine-tuning and evaluation
PythonPyTorchAWSSageMakerJupyter
Chapter 11 · DOCUMENT QA

Health Canada

Apr 2024 — Aug 2024
Ottawa, Canada
Machine Learning Engineer Intern
▶Built Llama3 7B + ChromaDB document QA for OECD report search and summarization
▶Lifted response and semantic accuracy ~20% via query transformation and contextual memory
PythonAzureLangChainStreamlit
Chapter 12 · OPERATIONS & AUTOMATION

Saputo

Jan 2024 — Apr 2024
Georgetown, Canada
Data Analyst Intern
▶Cut ~8 hrs/week of open-order updates with TypeScript Excel Office Scripts
▶Automated weekly workflows for 1,000+ Nestlé products, avoiding 20,000+ manual entries
▶Automated competitor research across 200+ brands with the Gemini API
▶Ran weekly EDA in VBA across 2,000+ products and 200+ miscellaneous SKUs
TypeScriptExcelVBAGemini API
Chapter 13 · RESUME PARSING

Respan AI

Y Combinator W24 · Formerly Keywords AI
Mar 2023 — Jun 2023
New York, United States
Software Engineer Intern
▶Parsed 1,000+ resumes with a spaCy NER pipeline to extract structured recruiter data
▶Cut parsing response delay 98% by adding SQLite result caching to the engine
PythonspaCySQLite
Chapter 14 · ENTITY RECOGNITION

Intapp

Formerly delphai
Jul 2022 — Sep 2022
Berlin, Germany
Machine Learning Engineer Intern
▶Raised entity recall 20% via improved labeling workflows and language-specific training sets
▶Hit 70% recall by fine-tuning spaCy NER and tuning hyperparameters with W&B on Azure
PythonspaCyW&BBS4
Operations · 作戦
Op · 01

OpenComposer

Mini-scale end-to-end reproduction of the Composer 2 training pipeline on a single NVIDIA GH200: MoE CPT, MTP self-distillation, Dr. GRPO + k1 KL, and behavioral auxiliary rewards.

PythonPyTorchMoEOpenRLHF
↗ REPO
Op · 02

AppGammon

Online backgammon with a React Native/Expo client, Hono backend, and shared gameplay rules in a TypeScript monorepo.

TypeScriptReact NativeExpoHono
↗ REPO
Op · 03

Ding-Bot

Chess engine combining GATEAU-style Graph Attention Networks with contrastive latent-space search

PythonTypeScriptGraph Neural Networks
↗ REPO↗ DEMO
Op · 04

PokerMon

Deep Counterfactual Regret Minimization (Deep CFR) for 6-player No-Limit Texas Hold'em

PythonDeep CFRGame Theory
↗ REPO↗ DEMO
Op · 05

LeaseEase

Streamlit app demystifying Canada’s Residential Tenancy Act with LLM + RAG, plain-language guidance, and auto-generated forms (T1, N7) for tenants navigating the housing crisis.

PythonStreamlitOpenAICohereChromaDB
↗ DEMO
Op · 06

MedChat

Assistant for clinical Q&A: Cohere Classify routes intent to a brain-tumor CNN or RAG over 1000+ WebMD pages with streamed answers in Streamlit.

PythonCohereTensorFlowStreamlit
↗ REPO
Op · 07

DirectU

Full-stack planner matching career goals and free-text course preferences to UWFlow reviews via Cohere, assembling a personalized four-year roadmap (React, Flask, MongoDB).

ReactFlaskMongoDBCohere
↗ DEMO
Op · 08

LeGM-Lab

AI-powered NBA take analyzer that fact-checks basketball opinions with real stats and roasts bad takes on X

PythonClaude APIFastAPIX API
↗ REPO↗ DEMO
Op · 09

FlightCal

Fetches flight info and exports it directly to Google Calendar or as an .ics file for any calendar app

TypeScriptNext.jsGoogle Calendar API
↗ REPO↗ DEMO
Stand Abilities · 能力
Languages
PythonSQLC++CTypeScriptJavaScriptRRacketBash
ML & Data
PyTorchCatBoostTensorFlowPySparkPandasNumPyGeoPandasspaCyMLflowHiveCUDA
Frameworks
FastAPIFlaskLangChainLangGraphReactNext.jsTailwind
Cloud & Databases
AWSAzureDockerGitPostgreSQLMongoDBSupabaseChromaDB
Contact · 接触

Your story has not ended.

— Connect with the Stand User

GitHubLinkedInEmailResumeXInstagram
CONFIDENTIAL · 極秘
〜 END OF RECORD · 記録終了 〜
秘