engagements
01–05
01
the shelved question
“we don’t want a body shop and we can’t get headcount until next year. build v1, hand my quants a model they can break, and tell us straight whether this dataset is worth anything.”
for a fund or proprietary trading desk, we take the dataset or model concept that has sat unevaluated, build the first version, and return an honest verdict as documented code your team keeps. the bayesian workflow behind it (Stan, PyMC, walk-forward validation, a posterior on whether the edge is real) is the reason the verdict holds up when your own quants attack it.
02
the statutory deadline
“we just got our LL97 letter and i don’t know if the number is right. how much are the fines going to be, and what’s cheapest to fix first?”
for a building owner or real estate portfolio, we deliver the scope 1/2/3 baseline that survives audit, the filing or GRESB submission itself, per-building fine exposure across scenarios, and a retrofit plan ordered by marginal cost per ton avoided. scenario modeling and uncertainty quantification are what turn a compliance spreadsheet into a defensible capital plan.
03
the efficacy claim
“we’re sitting on two years of wearable and blood data and our recommendations are still if-then rules in a spreadsheet. we can’t afford an RCT but investors want proof it works.”
for a consumer health or longevity company, we deliver an investor-grade outcomes analysis from the data you already have, honestly caveated, or design the causal recommendation engine itself. causal inference (the DAG, evidence tiering, dose-aware protocols) is what makes the claim survive diligence, and we have shipped exactly this: an engine covering 1,188 participants across 3 cohorts.
04
the filings pipeline
“every restructuring document is a PDF snowflake. we need one schema and a pipeline that survives all of them without an analyst re-keying each deal.”
for a company standing up AI workflows over doc-heavy legal and credit data, we design the canonical data model (holder, debtor, debt, position) and build LLM-assisted extraction with human review loops. uncertainty-aware confidence decides which documents route to a person, so the rigor shows up as analyst hours returned and a product your clients can query, not a methods lecture.
05
the no-comp number
“there’s no comp for this contract. how do we price it? i need a number i can defend to the investment committee, and if we’re wrong, i need to know which assumption we were wrong about.”
for a litigation-finance fund, a specialty insurer, or a corporate development team facing a one-off bet, we decompose the event into a sourced decision tree and propagate the uncertainty by simulation, so the committee sees not just the number but how wrong it could be and which assumption drives the error bar.