Cameron Olson

I turn ambiguous, badly‑measured problems into production models that survive regulatory review — and build the teams that run them.

Zero to one: ideating solutions to complex problems, then recruiting and leading the teams that tackle them. Twelve years of predictive systems with money and regulators attached — now building AI‑native measurement and decision infrastructure, with agentic tooling as the primary build method.

Founder Director / Head of

cameron@olson1.com LinkedIn ↗

B.A. Mathematics,
University of Hawai‘i at Mānoa.
Published in physics and
microbiology as an undergrad.

$310MDeployed on the national market-entry strategy my ranking model set.

93%Of two years' bookings driven by my target-customer model.

$24M+Annual NPV added while cutting outbound contact by half.

50MCustomers scored in production, under model risk governance.

$22MRaised and deployed to build a 100-space development from raw land.

Proof of work

Three that mattered

I Capital One 2013 – 2018

The model that became the national strategy

Formed the basis of our nationwide geographical strategy — by ideating, recruiting resources for, building a $200M business case around, and executing a Big Data Hadoop / Python / SQL machine learning empirical customer segmentation modeling and data infrastructure project, achieving VP and VP+ level approval and cross-functional support, ultimately managing 3 data scientists, 2 data analysts, and 1 GIS specialist. The results became the basis for our entire national geographical strategy (~$310M in spend to date), and were socialized with the CEO and Board of the company.

~$310M in spend to date Socialized to CEO and Board

II Capital One 2018 – 2020

Half the dials, $24M more NPV

Owned the Phone Contact Strategy agenda for the Card Loss Mitigation org, responsible for $300M+ in revenue and impacting 20M delinquent card customers annually, leading and mentoring 4 direct reports to successfully meet a Board commitment of 50% dialing reduction — cutting 480M annual dials to 20M customers by more than half, while actually growing NPV by $24M via analytically and test-determined customer and behavioral segmentations.

480M dials, cut by more than half $24M NPV growth

III AI startup (MVP) 2024 – 2025

An AI-first team, and the stack it shipped

Recruited cofounders and an AI-first development team, then led hands-on iteration of harnesses, sub-agents, and QC and auditing pipelines to build end-to-end scraping, ingestion, and tooling ML architecture — with an embedding + KNN retrieval and ranking stack, similarity clustering, and automated candidate scoring, in a PyTorch / Python + cloud + relational architecture — then built the endpoint tool and UI to serve the model results over tens of millions of assets.

Tens of millions of assets served Agentic tooling as the build method

02How I thinkFour selected essays

America has already differentiated into castes Census ACS, NHANES, PSID and BLS data on men 25–65 — India: 95% within-caste marriage; America: 93%. Nov 2025
624 likes · 167 comments

The “ten thousand PhDs” and what it tells us about AI risk and the next trillion dollar companies A SWOT on the datacenter-full-of-researchers framing, and where the value actually accrues. Mar 2025
62 likes · 13 comments

We can now gradient descent EVERYTHING Why cheap, scalable expert judgment collapses the old trade-off between legible metrics and real outcomes. Jul 2026
38 likes · 47 comments

3k? It is to laugh! Just how many billionaires are there in the world? A survey of legible and illegible wealth, and why the published counts are off by an order of magnitude. Nov 2025
42 likes · 4 comments

Writing at Performative Bafflement since 2025 — quantitative social science, AI, and economics. The top 5% of posts get 20,000–60,000 distinct readers.

03Hands onBuilt personally

~120-GPU compute installation Racks, power distribution, heat rejection, and daily uptime and margin monitoring — designed, assembled, and operated from scratch.

$50M production plant Dezhou, China. Bare floor to in production, on time and under budget, through a local plant manager and his organization.

1,200+ whp race car Built and instrumented personally: sensors, data logging, and tuning decisions made against logged data rather than feel.

A consumer product, and a company Invented and prototyped a flip-flop with a hidden pocket; grew Sandaloha to 7 employees and sales in 2 countries. Drafted the utility patent application myself — counsel cleared it for filing.

A 100-space RV park Financial model, $22M raise, then hands-on through sitework, utilities, and contractor oversight to an operating facility.

04CapabilitiesFounding · modeling · causal · governance · AI · engineering

Founding & zero to one

Standing up functions that did not exist · recruiting cofounders and first technical hires · founding, growing, and mentoring teams · taking an ambiguous problem from framing to business case to funded, production system · upskilling programs

Modeling

Loss forecasting and scenario analysis · propensity and risk scoring · gradient boosting (XGBoost, LightGBM) · random forests · logistic regression / GLM · segmentation and clustering · ranking and prioritization · embeddings and KNN retrieval · computer vision · anomaly detection

Causal & experimental

Matched-market synthetic baselining · quasi-experimental attribution · incrementality measurement · A/B and multi-arm design, power and sample sizing · multi-arm bandit optimization

Governance

Out-of-time validation · model risk governance and independent review · explainability and documentation for regulated use · drift monitoring and post-launch re-validation · credit risk governance submissions

AI & agentic systems

Agentic development workflows (Claude Code, Codex) as a primary build method · LLM and agent evaluation design, grading rubrics, inter-rater agreement · RLHF / RLVR trace methodology

Engineering

Python (pandas, scikit-learn, NumPy) · SQL, R, SAS · PyTorch, TensorFlow · Snowflake, Teradata, Spark, Hadoop, Databricks · AWS, Kubernetes · production feature and inference pipelines, API-served applications

If the problem is badly measured, that's the interesting part.

Best reached by email — I answer everything that isn't a template.

cameron@olson1.com

LinkedIn ↗