Chris Jones
Data Science & Analytics · New York City

I measure whether things actually work.

Senior BI Analyst at a solar fintech, where I founded the data science practice. I build recommenders, retrieval systems, and credit risk models, and I care most about how you prove any of them work. M.S. in Computer Science (machine learning) in progress at CU Boulder.

GitHub LinkedIn Tableau Email
Now Building a RAG system over SEC 10-K filings · Deep in the ML sequence at CU Boulder

Current focus

In development

RAG over SEC 10-K Filings

Python · embeddings · FAISS/Chroma · FastAPI

Question answering over 10-K filings for solar and clean energy companies. Ingestion and chunking strategy, vector retrieval, generation grounded in retrieved context, and an evaluation layer that grades retrieval precision and answer faithfulness rather than vibes.

Evidenceretrieval precision@k · answer faithfulness · grounded citations
Repo coming soon

Projects code + write-up for each

Sonic: Music Recommendation System

Python · EASE · implicit ALS · FastAPI

Implicit-feedback collaborative filtering on 1.68M Last.fm listening records. The evaluation harness came first: a leakage-safe per-user holdout with precision@k, recall@k, and NDCG written from scratch, frozen so the tuning loop couldn't edit its own grader, plus a sealed holdout read exactly once. A disciplined ALS sweep won on the small dataset and then lost on real data. Scaling up flipped the model ranking and put a linear item-item autoencoder (EASE) in production at NDCG@10 0.219 against a tuned ALS at 0.184, every gap significance-tested by paired bootstrap. Served through a containerized FastAPI service with a runtime diversity lever and an explicit fallback strategy field. Deployed, with the 514 MiB model matrix fetched and checksum-verified at boot instead of refitted.

Evidencefrozen eval layer · sealed holdout · paired bootstrap CIs · strong baselines, not just popularity · calibrated against published SOTA
Try it live → · API → · Code on GitHub → · Read the deep dive →

Credit Card Fraud Detection

Python · scikit-learn · Streamlit

Six models on 492 frauds hidden in 284,807 transactions. Anomaly detection sounds like the natural fit for fraud and loses badly: Random Forest reaches 0.96 precision at 0.76 recall (F1 0.85) while the best unsupervised model manages F1 0.22. Accuracy is excluded from the scorecard on purpose, because predicting "not fraud" every time scores 99.83%. Ships with a threshold-adjustable dashboard, because where to sit on the precision-recall curve is a business decision, not a modeling one.

Evidenceprecision-recall focus · 6-model comparison · accuracy deliberately excluded · threshold tuning
Code → · Bake-off →

Telco Customer Clustering

Python · scikit-learn · PCA · Streamlit

K-Means over 7,032 telecom customers, with the cluster count picked by a silhouette sweep rather than by squinting at an elbow plot, PCA for checking structure, and per-cluster churn profiling to make the segments mean something: churn ranges from 4% to 59% across the ten groups, which is the difference between a coordinate and a retention conversation. Dashboard for filtering and exporting segments.

Evidencesilhouette sweep · PCA structure checks · per-segment churn profiling
Code → · Field notes →

Writing

Experience 2020 - present

Senior Business Intelligence Analyst

Sungage Financial · Feb 2026 - present

Lead analytics for the Credit, Revenue, Product, and Strategic Markets teams, and founded the data science practice that now sets the company’s predictive modeling roadmap. Promoted on the strength of the causal inference and credit risk work below.

Business Intelligence Analyst

Sungage Financial · Feb 2024 - Feb 2026

Built the Account Health monitoring framework the revenue organization runs on daily, covering several hundred partner accounts through a period of rapid growth. Delivered the uplift analysis on a sales incentive program against a large control population and recommended against scaling it, the answer leadership needed rather than the one it wanted.

Associate Business Intelligence Analyst

Sungage Financial · Feb 2023 - Feb 2024

Joined as the analytics function was forming and supported Revenue, Marketing, and Strategic Markets through the company's first major growth inflection. Promoted within twelve months.

Environmental Field Data Scientist

BCA Environmental Consultants · Jan 2022 - Feb 2023

Analyzed soil, water, and demolition datasets underpinning regulatory compliance filings and client-facing technical deliverables, on projects where the defensibility of a result rests as much on collection conditions as on the analysis applied to it.

Biological Science Aid

USDA Agricultural Research Service · Feb 2020 - Dec 2021

Research assistant on the study behind the first report of Fusarium fujikuroi causing root rot and seedling elongation in Indiana soybean, supporting the lead researcher across inoculation trials, data collection, and analysis. Second author on the resulting note in Plant Disease 105(11), published during undergraduate study at Purdue.

Selected Professional Work

Sales Initiative Causal Inference Framework

Python · scikit-learn, LightGBM, XGBoost, scipy

Reusable framework answering whether an initiative caused a change or would have happened anyway, using propensity score matching, balance diagnostics, and Bonferroni-corrected testing. First production run evaluated a sales incentive program against a large control population.

Code and details confidential

First Payment Delinquency Model

Python · LightGBM, SHAP, pandas

The company's first production ML project: a calibrated delinquency model for residential solar loans, built on point-in-time feature engineering, time-based vintage splits, and leakage prevention by design. Its five-phase project framework is now the internal template for data science work.

Code and details confidential

About

I came to data the long way around: soil science at Purdue, plant pathology research with the USDA, then environmental field work, where I learned that messy real-world data is the only kind there is. Science training is where the habit comes from. In a lab you design the experiment before you run it, because a result you can rearrange after the fact isn't a result. Since 2023 I've been at Sungage Financial, promoted twice as the company grew quickly, and I founded the data science working group along the way.

The habit that runs through everything here: decide how you'll know it worked before you build it. Which mostly means being willing to find out that it didn't. The recommender write-up spends a section on the moment a well-tested conclusion turned out to be true of my dataset and false of the problem, and the fraud bake-off exists because the obvious tool for the job lost. Those are the parts worth reading.

Off the clock I'm usually on a bike, or somewhere in the Lands Between.