Quantitative analytics for energy and finance
Grid, price, and valuation models that hold up in the room.
Independent quantitative analysis for energy infrastructure and finance teams. Every engagement ends with a documented model and a running application, not a slide deck.
I reply within one business day. Based in Miami, Florida, working with US and EU teams.
PJM zone vulnerability, baseline
21 zones ranked under data center load growth
Live baseline from the ComputePower model. Hover or tap a zone. The three Mid-Atlantic zones hosting the Northern Virginia, DC, and Baltimore cluster score roughly three times any other zone.
Who this is for
Teams that have to defend a number.
The work is for people who present analysis to a committee, a board, or a regulator, and who need the method to survive the second question.
Not the right fit for
- Business intelligence dashboards and routine reporting
- Staff augmentation or hourly data engineering
- Personal investment advice of any kind
Infrastructure and siting teams
Data center developers, hyperscaler energy teams, and infrastructure funds deciding where new load can be built.
“Which zones absorb another 500 MW without volatility blowing out, and how far does queued generation actually relieve them?”
Power market analysts and risk desks
Analysts and risk managers at utilities, IPPs, and trading desks who need to explain price behavior, not only report it.
“How much of a zone’s LMP volatility is load growth, and what does it look like under a scenario we have not seen yet?”
Investment and valuation teams
Corporate development, private capital, and research teams who need a number they can defend in committee.
“Is the figure on the term sheet consistent with the data, and which assumption moves it most?”
What an engagement produces
Four things, every time.
01
A ranked answer with the assumptions visible
Not a heat map. A ranking, a score, and the components that produced it, so the person presenting it can explain why zone three is third.
02
A model that re-runs when the inputs move
Scenario controls and fixed normalization, so next quarter’s question is a re-run, not a new project.
03
A written result that states its limits
What the method covers, what it extrapolates, and where it should not be used. Committees trust the analysis that names its own edge.
04
A deployed application, not a notebook
Maps, charts, and an API for the people who will use the result and would never open the code.
Selected work
ComputePower: PJM Grid Vulnerability Under Data Center Load Growth
Data center electricity demand in the United States nearly doubled between 2023 and 2026, and grid capacity has become the binding constraint on where new compute can be sited. Ranking which zones absorb that load safely means holding several moving quantities together at once: price volatility, congestion stress, existing data center concentration, and the queued generation that might relieve any of it.
Result. All 21 PJM zones ranked and re-ranked in real time as load growth and queued solar assumptions move, presented on an interactive map for readers who will not open a notebook. Because normalization anchors are fixed to the baseline period, a zone's score means the same thing across scenarios, which is what makes the comparison usable rather than merely responsive.
Out-of-sample R² of 0.163 on 2024 to 2026 data, stated in the case study because a screening tool that hides its fit is not one.
- 21
- PJM zones ranked
- ~1 GB
- hourly market and load data
- 3
- independent methods
- 608
- data center facilities mapped
systems
Blackletter: Legal Document Intelligence
A retrieval-augmented system for interpreting statutes, contracts, and procedural material, built so every output stays traceable to a source.
systems
Olist: E-Commerce Marketplace Analytics
A migration and analytics system over Brazilian marketplace data, including a geospatial application mapping 35,000 seller-buyer trade routes across 1,600 cities.
Engagements
Start small, then decide.
01
Scoping diagnostic
2 weeks
Teams with a question and uncertain data, before committing to a larger project.
What you get →02
Analysis engagement
4 to 8 weeks
Teams that need a defined question answered with a documented model.
What you get →03
Model to application
6 to 12 weeks
Teams whose model has to be used by people who will not open a notebook.
What you get →04
Retained advisory
Monthly, 3-month minimum
In-house analytics teams that want an independent review of method and interpretation.
What you get →
How an engagement runs
The decision first, the data second, the model third.
Step 1
Scope the decision
Start from the decision the analysis has to support and the person who will make it. The question is written down and agreed before any data is touched.
Step 2
Audit the data
Establish what the data can and cannot support. Coverage, granularity, gaps, and the difference between what a field is called and what it measures.
Step 3
Model with more than one method
Each method answers a different objection. A machine learning model, a panel regression, and a causal design agreeing is a stronger claim than one well-fit model.
Step 4
Validate and state the limits
Out-of-sample tests, fixed normalization so scenarios stay comparable, and a plain statement of what the result does not cover.
Step 5
Deliver as software and a written result
A memo for the committee and a deployed application for the team. The reasoning is inspectable rather than trusted.
Background
Economics first. Then twenty years of systems that fail in public.
M.A. in Finance and Banking and B.A. in Economics from Charles University, both with honors, then valuation and M&A work at Erste Group, including the tender of a real estate portfolio that closed at $160 million.
Two decades running operations at CNN, NBCUniversal, and Czech Television followed: a financial audit and cost model for CNN, election night graphics for an audience above 18 million at NBC, and a business news channel launched on a $20 million investment. Live broadcast is an unusual place to learn systems design and a clarifying one. Every workflow either holds under load or breaks in front of an audience.
That is the standard the models are built to. More about the background or the full CV.
Next step
Bring the question. A 30-minute call decides whether it is answerable.
Describe the decision, the data you have, and when you need an answer. I reply within one business day.