Selected work

Engagements,
told plainly.

A handful of recent projects across data platforms, decision systems and LLM products. Some clients prefer to stay anonymous; the numbers and the shape of the work are real.

Financial services · 2025
Series B fintech
Data ScienceForecastingProduct

Forecasting platform that the CFO actually trusts

Challenge
Revenue forecasts were rebuilt by hand each month in a sprawling spreadsheet. Finance and the data team disagreed on the numbers.
Approach
Designed a forecasting service with reproducible models, scenario inputs and a thin review UI. Wrote evals against past months so accuracy is measured, not asserted.
Outcome
Monthly close cut from 6 days to 2. Forecast error down ~38%. One source of truth shared by Finance, RevOps and the board deck.
Manufacturing · 2024
Industrial SaaS scale-up
Data EngineeringdbtDatabricks

Lakehouse rebuild without downtime

Challenge
Legacy warehouse was slow, expensive and undocumented. Every dashboard had a different definition of 'active customer'.
Approach
Stood up a Databricks lakehouse with dbt models, contracts and lineage. Migrated workloads in slices, with shadow runs to verify parity before cut-over.
Outcome
Compute spend down 54%. Median query 12× faster. Definitions for the top 30 metrics now versioned and owned.
Health · regulated · 2025
Healthcare platform
LLM SystemsEvalsHealthcare

LLM clinical-notes assistant with guardrails

Challenge
Clinicians spent 90+ minutes a day on documentation. Off-the-shelf LLM tools couldn't meet the safety bar.
Approach
Designed a retrieval-grounded assistant with structured outputs, PII redaction, a domain eval set and human-in-the-loop review for high-risk fields.
Outcome
Pilot with 40 clinicians cut documentation time by 47%. Zero unredacted PII incidents in 90 days of use.
Consumer marketplace · 2024
Marketplace, Series C
ExperimentationCausalProduct

Experimentation platform for a small data team

Challenge
Product teams shipped fast but couldn't tell which changes mattered. A/B tests were ad-hoc and often misread.
Approach
Built an experimentation service on top of the warehouse: assignment, variance reduction, sequential tests, and a results UI product managers could actually read.
Outcome
Test velocity 3× in one quarter. Three previously-shipped 'wins' caught as flat or negative and rolled back.
Supply chain · 2023
Logistics carrier
Decision SystemsOperations Research

Routing decisions, made by the system

Challenge
Dispatchers were the bottleneck. Routing decisions lived in their heads and didn't scale into new regions.
Approach
Co-designed a decision system that proposes routes, surfaces tradeoffs, and learns from dispatcher overrides. Shipped behind a feature-flagged rollout.
Outcome
On-time delivery up 6 points. Dispatchers handle 2.4× more volume per shift. Expansion into two new regions unblocked.

Have a project that belongs here?

A 30-minute call, no decks. We'll tell you honestly whether we're the right team for it.

Start a conversation