MSc Finance (STEM-designated), Hult International Business School · 2026 BBA, Finance & IT, HEC Montréal · 2024
RECOGNITION
01x 1st place, Personal Agent Hackathon, NYU Tandon (Coverage Brain) 01x 2nd place, Ethical Tech CoLab hackathon, hosted by Microsoft Research Lab 01x Feature design & analysis, Matatah (pre-seed round in progress) 18x Case studies with source files attached 04x Live apps
Finance / analytics / AI-enabled research
Modeling and valuing energy and infrastructure assets, and building the practical digital tools behind them. Finance × IT, from spreadsheet to shipped app.
I’m Alexandre Ly. This is a working portfolio of energy and markets research, valuation models, AI strategy cases, and machine-learning studies, plus four live apps, each built to make assumptions visible and conclusions defensible.
Where each case was made, and what it studied. Pick a case to trace its line.
Flight route
Made in → studied
◆Subject location
Drag to spin
02 / About
The person behind the models.
I’m a finance analyst with an IT background: BBA in Finance & IT from HEC Montréal, MSc Finance from Hult. Before the MSc I worked inside a utility and a bank, coordinating cybersecurity solutions at Hydro-Québec and then implementing a banking platform at National Bank Financial. That is where I picked up the habit of treating a model as a system someone has to operate, not a course deliverable. My research focus now is how AI-driven load growth changes utility capital, rates, and valuation. Bilingual, French and English.
Professional experience
Banking Platform Implementation SpecialistNational Bank Financial
Operations Solutions AnalystClassic Woodwork
Cybersecurity Solutions CoordinatorHydro-Québec
03 / Selected work
Twenty ways to move from a spreadsheet to a point of view.
The source files stay attached to each case study. Start with the decision, then open the report, model, or presentation when you want to go deeper.
Pick an area, then a project. Its result opens on the right.
01Load forecasting / Utility planningMade in New York City
Dominion Energy load forecast
Probabilistic data center demand, 2025–2035, and what it does to capital and rates
A probabilistic forecast of data center demand in Dominion Energy Virginia’s territory (PJM DOM zone). A weather-normalized baseline and a stage-weighted pipeline of 46 projects feed a Monte Carlo simulation, which is then carried through capacity, capital, rate base, customer rates and utility earnings, with a stranded-load stress test and a large-load tariff.
Model chart / new data center peak, 90% range13.0 GW by 2035
90% range (P5–P95)median (P50)
12,957 MWmedian new peak, 2035
$83.4Bbase-case capital
+10.1%base-case 2035 rate effect
90% range 8,661–16,569 MW · 46 projects, 17,530 MW · peak model MAPE 5.05% vs 14.33% naïve · stranded load: +13.8% rates, +5.3% with a large-load tariff
02Equity researchMade in Miami
NextEra Energy
Five-scenario renewable utility research
A full initiation report on the regulated utility and renewable platform, pairing peer-derived WACC, a five-scenario DCF, EV/EBITDA cross-checks, and an AI-energy demand thesis.
Model chart / five scenariosbase 6.8% CAGR
baseupside / managementdownside
$9712-month target
5.75%WACC (base)
BUYrating
Base DCF: $95.58 · 12-month target: $97 · management case: $104+ · report + model
03Portfolio construction / IPSMade in Vancouver
AI–Energy Nexus
ACWI active portfolio construction
A Treynor–Black active overlay on ACWI (iShares MSCI ACWI ETF) using seven grid and energy-infrastructure names, translated into a client-specific capital allocation and monitoring dashboard, and governed by an Investment Policy Statement.
Model chart / complete portfolio mix18.61% expected return
An electrification investment case: a five-year model for electric mini concrete mixers in Naples’ restricted traffic zones, linking unit economics, labor, permits, tax incentives, and free cash flow to a project NPV.
Model chart / FCFF profile€264k NPV
initial capexannual FCFF€k · 8.5% WACC
€264kproject NPV
€390kinitial capex
€891kyear-one sales
Five-year model · WACC 8.5% · Italy’s Transizione 5.0 credit and ZES Unica zone
05Tariff modeling / Data center powerMade in New York City
GRID / OHIO
What a 100 MW data center pays under AEP Ohio’s 2025 tariff, and what a contract, a market or a flexible load changes
An independent, interactive cost-to-serve study of a hypothetical 100 MW data center under Ohio Power Company’s (AEP Ohio) data center tariff, ordered in July 2025. It rebuilds the minimum-demand bill, tests a partial 70 MW fixed-price block, stress-tests gas, capacity and flexible load, and compares PJM and MISO energy prices over the same 8,760 hours.
Model chart / PJM cost of a +30% gas shock$3.75M to $0.83M
unhedgedblock plus indexadded annual cost, 2025 hours
$2.64Ma year, minimum-demand charge
77.8%of energy covered by a 70 MW block
$1.09Mlower energy cost in MISO
Sole author · historical 2025 model · 788,400 MWh · PJM AEP-Dayton $45.74/MWh vs MISO Indiana $44.36/MWh · +50% capacity price adds $2.74M a year · 54 sensitivity cases, 10,000-trial Monte Carlo
06Gas offer pricing / Natural gas supplyMade in New York City
Gas Offer Planner
From market inputs to a clearer natural gas offer for commercial and industrial customers
A guided, supplier-side planner for pricing a natural gas offer. Six steps take a customer, an annual gas quantity and weather assumption, a dated market snapshot (Henry Hub plus regional basis), the full supplier cost stack and a credit check through to a review, with a customer proposal and an internal financial review as anonymized PDFs. Anything not yet known stays pending, and every calculation runs on the visitor’s device.
Model chart / the six stepscustomer to PDF
6 stepsfrom customer to reviewed offer
6,100MMBtu a year, the estimated benchmark
2 PDFscustomer proposal and internal review
Sole author · estimated operating profile · Tulsa, Omaha or Columbus, Ohio · 12 or 24 months from January 2027 · bundled Tulsa 2027 prices dated October 29, 2026, indicative, not an executable quote · no trades, no live market requests
07AI strategy / TransformationMade in San Francisco
Relay Pulse
An AI transformation strategy for an activewear brand
A team strategy for a company whose December demand shift went unnoticed for eleven days. It diagnoses the failure with the Viable System Model, then recommends a low-cost alarm layer (Relay Pulse) instead of a $2.8M personalization engine, with a governance rail, a three-KPI panel, and a phased roadmap.
Model chart / vanity–viable verdicts3 use cases
VantagePoint engineRefuse as scoped$2.8M over 3 years · metrics can rise while margin falls
Relay PulseBuildowned data · weeks to pilot · internal only
CX CopilotPilot, phase 2a human sends every message · only after Pulse proves governance
Verdicts from the team's Vanity–Viable Matrix · Relay is a course case company
14hmedian signal-to-action
81%alerts acted on within SLA
$11.08Mest. revenue protected
Team project · VSM, THRIVE, and CLAIM frameworks · figures come from the team's prototype dashboard and December replay
08AI strategy / ReadinessMade in Singapore
AI Readiness Assessment
A Canadian bank's Copilot pilot, scored on eight dimensions
A readiness assessment, delivered as a podcast script, of a major Canadian bank's Microsoft Copilot pilot. One anonymous interview with a banking analyst is scored on the eight dimensions of Tehrani et al. (2024), ending in an adoption strategy and five recommendations.
Model chart / readiness scorecardready to start, not to scale
Dimension
Not evidenced
Partial
Strong
Informational
Technological
Environmental
Infrastructural
Data
Participants
Customers
Process
strongpartialnot evidenced
Ratings from the assessment's Appendix A · one interview, so “not evidenced” means the topic did not come up
2 of 8rated strong
5 of 8partly ready
8dimensions scored
Individual assignment, Agentic AI for Business Transformation coursework · verdict: a controlled, human-reviewed pilot
09AI implementation / LocalizationMade in Singapore
Netflix AI Localization
Simplify-then-translate, and whether Netflix was ready for it
A team case study of how Netflix moved subtitle localization from human translators and vendors toward AI. It audits readiness across culture, data, leadership, and talent, scores the use case against eight agent-readiness criteria, and lays out a six-step implementation that keeps humans in the loop.
Model chart / agent-readiness scorecard6 of 8 criteria met
Criterion
Fit
Partial
Specialization bottleneck
24/7 availability
Repetitive or tedious
Well-defined process
Measurable
Error tolerance
Complex decision
Personalization
fitpartial
Ratings as scored on the deck's readiness slide · partial scores are why a human finalizes every subtitle
6 of 8criteria met
9.5Mlabeled subtitle pairs
6implementation steps
Team project · 2026 · built from public sources, with no savings figures · team video included
10AI in finance / Agentic RAGMade in San Francisco
Agentic Trading Signal
News to BUY / HOLD / SELL, with cited reasoning
A team-built agent that turns a ticker into a trading signal: Yahoo Finance news is embedded with ModernBERT, indexed in MongoDB Atlas vector search, and read by Gemma 3, which returns a signal, a confidence level and its reasoning. Positioned as research triage; not backtested.
Model chart / RAG pipelineModernBERT → Gemma 3
Ingest
Yahoo Finance10 news items per ticker
Chunk500-character pieces
ModernBERT768-d embeddings
MongoDB Atlasvector index, cosine
Query
Ticker“what news suggests it may rise or fall?”
ModernBERTquery vector, 768-d
Vector searchtop 5 articles retrieved
Gemma 3 1Bsignal · confidence · reasoning
After the ticker is entered, retrieval and generation run without human input · the analyst reviews the output
768embedding dimensions
Top 5articles retrieved per query
10news items per ticker
Team project, AI Trends in Finance coursework · Python · Gradio · TSLA notebook run: HOLD at medium confidence
11FinTech / PaymentsMade in Boston
Stripe
Is Stripe a FinTech, what does FedNow do to it, and how should its teams collaborate?
A team presentation that tests Stripe against five FinTech criteria, contrasts its developer-first model with legacy merchant acquirers, then asks how the Federal Reserve's FedNow rail could help or hurt it. At its core is a seven-step business model of one Stripe payment, from merchant onboarding to payout.
A separate six-author group paper looks at Stripe from the inside and asks where collaboration breaks: speed against safety, global against local, and siloed incentives between teams such as Payments and Billing. Its design shifts coordination upstream with three principles and three mechanisms, and it is explicit about what it accepts in return: less voice in some decisions, uneven collaboration quality, and more information for teams to process.
Model chart / settlement speed by rail3 of 5 rails shown
Card networksDelayedinstant authorization · settlement follows · about 2–3% in fees
ACH1–3 daysFed + The Clearing House · traditional payment rail
Speeds from the deck's rail comparison · wallets and blockchain were also covered and are left out here
Model chart / collaboration design3 failures → 3 mechanisms
Where collaboration breaks
Speed vs safetycompliance debt, and violations found after launch
Global vs localregional squads retrofit compliance, as with Brazil
Siloed incentivesPayments and Billing built incompatible APIs
The design's answer
Structure firstcross-functional teams catch compliance while a feature is built
Shared global coreglobal APIs and design docs, adapted by regional squads
Aligned incentivesjoint decision forums weigh growth and risk
Mechanisms
Design Dossierone living record of core payment architecture
Ownership Mapsa global owner and local owners for each critical component
Risk Escalationa cross-functional chain for regulatory and financial risk
From the collaboration design paper · principles and mechanisms as the paper names them
7steps in the payment model
4FedNow risks flagged
3collaboration mechanisms
Team deck, six members · separate collaboration-design paper, six authors · qualitative analysis, no valuation figures
12Investment case / Sustainable financeMade in Boston
J.P. Morgan Efficiency Swap
A sustainability-linked swap that prices data center energy efficiency
An investment case for J.P. Morgan, built for Financial Markets & Sustainability coursework. It proposes an interest rate swap that lowers the spread U.S. data center developers pay when they reach an efficiency target, and sets out the stakeholder pressure, the balance-sheet logic, the verification standards against greenwashing and a two-phase pilot in Northern Virginia.
Model chart / the Efficiency SwapPUE-linked pricing
Step-down−10 bpPUE of 1.2 or lower · a lower cost of capital for efficient operators
Step-upPenaltyPUE target missed · size not stated in the deck
VerificationBig FourDeloitte, PwC, EY or KPMG verify PUE and carbon data first
Terms as stated in the deck · PUE is Power Usage Effectiveness, the standard data center efficiency KPI
10 bpspread step-down at PUE ≤ 1.2
PUE 1.2efficiency threshold for the step-down
$2.5TJPM 2030 sustainable-finance target
Financial Markets & Sustainability coursework · figures as stated in the deck, not independently verified
13Investment case / AI FinTechMade in Boston
Project Meridian
An investment case for an AI risk-intelligence company
A team investment case for an AI company that sells risk intelligence to family offices. It sizes Asia-Pacific and UK wealth, tests family-office demand, prices against Bloomberg and AQUMON, and sequences fundraising so the product earns credibility before it raises equity.
Model chart / annual running cost by path4x cheaper structurally
fintech path (chosen)asset management (deferred)
Annual cost from the deck's comparison · asset management also needs seed capital
$99TAPAC wealth by 2029
~4,000HK and SG family offices
80%+UK advisors on model portfolios
Team investment challenge · “tech first, fund later” · market figures as cited in the deck · no round size or valuation
14Quantitative modeling / RegressionMade in San Francisco
Diamond Price Regression
Interpretable log–log OLS, plus a re-check of its R²
An OLS study on 53,940 diamonds that treats interpretability as a constraint: one size variable instead of three collinear ones, ordinal quality grades, and log–log elasticities. The case study then tests what the headline R² actually measures.
Model chart / R² by specification0.997 → 0.838 centered
no interceptwith intercept○ as reported
R² on log price, centered basis · hollow marker is the uncentered value statsmodels reports when the intercept is dropped
0.838centered R² · final spec
0.950same spec + intercept
96.7%of rows reconcile on depth %
53,940 rows · condition number 10.7 vs 97.7 · Python · statsmodels · Phik · hold-out MAPE on price 33% (with intercept) vs 13% (no intercept)
15Machine learning / PredictionMade in San Francisco
Apprentice Chef
Revenue and cross-sell models, judged on the train–test gap
Two prediction studies on a meal-kit business: a revenue regression and a cross-sell classifier, each ending in a weighted blend of models. The case study compares every model out of sample and checks which results survive a re-run.
Model chart / train → test AUCcross-sell classifier
logisticfinal blend○ train · ● test
Test AUC by model · final blend re-run gave 0.890 → 0.712 (gap 0.179)
0.726logistic test AUC
0.777revenue blend test R²
80% vs 42%subscribe rate, professional vs consumer email domains
1,946 customers · 75/25 split · Optuna tuning · the cross-sell blend's gap depends on an unseeded tuner
16Data analytics / Portfolio reviewMade in Boston
Client Portfolio Analytics
SQL extraction and a Tableau dashboard for a client portfolio review
An individual assignment that rebuilds one client's portfolio from a wealth-management database with SQL, measures returns, volatility, and return-to-volatility ratios by security, and turns them into a Tableau dashboard and a management report that flags concentration risk and recommends a rebalance.
Model chart / Return / volatility vs position sizelargest holdings, lowest return / vol
alternative ETFslarge equity positions
Return-to-volatility (Sharpe at rf = 0) from the Tableau dashboard, position value from its position chart
61%of position value in four holdings
0.1539best return / vol (CTA)
15%suggested trim of ROK, BIDU
Individual assignment, Data Extraction & Visualization coursework · SQL, Tableau · daily returns as of 9 Sep 2022
17Valuation / DCFMade in Singapore
Grab Holdings
DCF + public-company cross-check
A USD valuation of the Southeast Asian superapp, combining an integrated DCF with guideline public-company analysis and a thesis around profitability inflection, network effects, and net cash.
Model chart / revenue ramp$11.1B FY35E
Revenue in USD · historical 2021–25, forecast 2026–35
$4.52base fair value
+29%implied upside
BUYrecommendation
Sensitivity range: $3.74–5.67 · WACC 9.91% · PDF + Excel model
18Valuation / DCFMade in Boston
Planet Fitness
Recurring-revenue DCF
A DCF for a franchised fitness platform, built around same-store sales, margin progression, capex, working capital, terminal value, and exit-multiple sensitivity.
Model chart / DCF build$1.66B revenue 2029E
revenueEBITDA$mm · base case
$82.34implied share price
10.28%WACC
18.0xexit multiple
Modeled 2021–2029 · team presentation included · Excel + MP4 · peer data from PitchBook
19Equity research / DCF valuationMade in Singapore
Netflix Equity Research
A BUY pitch: a $115 DCF target against a ~$75 analyst view
A three-analyst investment committee pitch on Netflix. It argues that the market and standard analyst models underprice the company's operating leverage, builds a ten-year DCF at an 8.2% WACC to reach a $115 target (+31% on $87.69), and cross-checks it against peer multiples. A supplemental workbook adds peer and financial statement analysis.
Model chart / where the target sits$75 · $87.69 · $115
Prices from the pitch · the ~$75 analyst value is the pitch's stand-in for standard analyst models
$115DCF target, +31% on $87.69
8.2%WACC, vs 8.9% analyst case
8.5%10-yr revenue CAGR, vs 4.5%
Team of three · investment committee pitch, 14 May 2026 · ten-year DCF, EV bridge and peer multiples cross-check
20Financial analysis / Variance modelMade in Boston
Walmart Financial Analysis
Three years of 10-K results, and how well Walmart's guidance and analyst forecasts held
A team analysis of Walmart's fiscal 2023–2025 results: statement and ratio analysis, a variance model that sets actuals against Morningstar forecasts and management guidance, and an evaluation of Walmart's budgeting and forecasting process with seven proposed improvements.
Model chart / net income vs analyst forecast+17.0% → −1.0%
above forecastbelow forecast
Actual net income vs Morningstar's forecast, from the variance model · the surprise shrank each year
+66%net income, FY2023–25
22.2%FY2025 ROE, up from 14.6%
−0.25%FY2025 revenue vs forecast
Team project, Cost & Managerial Analysis coursework · public 10-K data, Morningstar forecasts, and earnings-call guidance
04 / Apps
Apps I’ve built.
Working products, each with a live link so they can be tried rather than described.
01
Matatah
A hiring platform with a two-sided trust feature
AI product / Hiring trustMade in New York City
GoalHelp candidates and recruiters trust each other, with evidence instead of a score.
A hiring platform that shows the evidence behind trust to both candidates and recruiters. My part: feature design and analysis of the trust feature.
My part Feature design and analysis of the trust feature
Context Hackathon team project · Trust in the Hiring Funnel Hackathon
A supply-chain explorer for NVIDIA’s mineral dependencies
Supply chain / ResilienceMade in New York City
GoalShow how a mineral disruption reaches NVIDIA, and what resilience is worth.
Explore NVIDIA’s disclosed suppliers and mineral facilities, and stress-test how a disruption travels. 2nd place, Ethical Tech CoLab.
My part Coordinated the Resilience Index feature, and made the case that Mineral Trace should be more than a monitoring app: a solution giver and a stress-test calculator as well
An AI notetaker that turns a meeting into owned action items, decisions and risks, each traced to what was said
AI agents / Meeting operationsMade in San Francisco
GoalTurn a meeting into follow-through: who does what by when, with the exact words behind every item.
It takes a meeting three ways: a pasted transcript or recording, a live call, or a bot sent into a Google Meet, Teams, or Zoom link. faster-whisper transcribes it; the app then extracts a summary, decisions, owned action items, risks, and follow-up drafts, each with a verbatim quote. Slack threads and calendar reminders are drafted for review; nothing goes out until a person approves it.
My part Set up and configured faster-whisper for transcription and pushed it to the repository
Rule Every item is traced to what was said, and a person reviews before anything is sent
Stack faster-whisper transcription, a language-model extraction step, GMI Cloud
One day of electricity demand on California’s grid (CAISO). Scroll to move the clock, or drag the time slider, and add a data center to see what it does to the evening.
Open to 2026 analyst roles in energy and infrastructure finance, utility planning, and equity research, and to conversations about the decisions that make a model useful outside the spreadsheet.
