Systems I run at work, things I build on my own time, and where I started. Each card says what it does, what came out of it, and the stack. Nothing here is a mockup.
The internal systems a commercial solar and energy-storage contractor runs its back office on. I'm the sole developer. The code is private, so these describe what each system does and how it's designed, with no company data. The lessons go deeper.
The accounting system stays authoritative. Everything to its right is a derived, dated view — it can explain the truth, but it never gets to outrank the source.
Work · sole developer
A Python + SQLite platform that consolidates accounting, banking, project, and procurement data into normalized tables, exception queues, and generated reports. A strict source-of-truth hierarchy keeps the accounting system authoritative; everything else is a derived, dated view.
Work
Generates QuickBooks Enterprise import files instead of hand-keying transactions. Looks up what already exists, checks for duplicates, enforces field limits, validates sources, and fails closed. A person always performs the final import.
Work
Matches bank activity against accounting and project records using normalized identifiers and confidence scoring. Every record lands in an explicit state (confirmed, proposed, unmatched, or needs review) instead of being forced into a match.
Work
A warehouse health view, an agent-coordination dashboard, and an ownership board that shows every open item's owner and how long it's been waiting. All generated from the same pipeline, so they can't drift from each other the way one-off dashboards do.
Work
Reads incoming PDFs, including scanned ones. Uses OCR only when native text isn't enough, extracts structured fields into a review step, preserves existing data, and fails loudly instead of treating an unreadable file as empty.
Work · evolving
Specialized Claude Code agents with task handoffs, recorded decisions, and verification rules. A release gate and a growing automated test suite check every change before it ships, not just the parts that touch money. Started bigger; I cut it down to what actually earned its coordination cost.
The part worth seeing, not just reading about. Three features built without a framework: a live countdown and cast carousel, a multi-source news feed, and a release tracker grouped by date confidence.
Things I build to learn, and keep using.
Personal data lakehouse
A local-first data lakehouse that turns my scattered personal exports into structured, searchable data. Raw files land untouched, Python parsers normalize them into SQLite with full-text search, and curated, git-versioned records feed generated views for each AI tool. Search and summarization run on local models, and an explicit boundary rule decides what may ever reach a cloud model.
Public proof-of-concept
The matching pattern from my work reconciliation engine, rebuilt from scratch on fake data so it's safe to publish: confidence-scored matching, duplicate detection, and four explicit states (confirmed, proposed, needs review, unmatched) instead of forcing a match. Synthetic transactions only — nothing here is the real code.
Live app
A dairy-ingredient checker that returns SAFE, UNSAFE, or REVIEW REQUIRED with a plain-language explanation. It catches what labels hide: casein in "non-dairy" products, milk-derived additives, and vague terms. Rules live in YAML so the logic stays auditable; product lookups enrich the result.
SavvyCoders capstone · still maintained
A movie, TV, and gaming discovery app built from scratch with no framework: native ES modules, a hand-rolled component layer, client-side routing, and an Express + MongoDB API over TMDB data. I've kept shipping to it since graduation, including security hardening for untrusted API text.
Personal project
A Midwest angler's map and field guide: 200+ curated lakes and rivers, a species catalog, a live overlay of Illinois lake-depth data from the state's ArcGIS service, and boat ramps and piers from OpenStreetMap with fallback across mirrors. No backend; state lives in the browser.
Personal project
My garden as a data system: registries for areas, zones, and plants, observation and harvest logs, weather inputs, a generated daily hub, and a phone-friendly form backed by a small local server that writes observations back into the dataset.
Personal project
Image and language models running entirely on a 12 GB laptop GPU: LoRA training with checkpoint-by-checkpoint evaluation, a ComfyUI pipeline with face correction and 4x upscaling, and 30B-class language models on llama.cpp. Most of the work was measuring and fitting into memory, not picking models.
A hands-on solar lab in my garage for learning DC power, MPPT, sensing, and telemetry from first principles. Every component decision is logged with its datasheet math, and nothing gets bought until the measurements exist.
ESP32 · MicroPython · INA228 · solar PV
The next step: putting a real computer workload behind a small solar + battery system to explore telemetry, load shedding, and energy-aware computing at miniature data-center scale.
Salvaged PC hardware · ESP32 · Python · SQLite
16 end-to-end projects from the TripleTen program: churn, pricing, forecasting, NLP, and computer vision. Highlights: telecom churn with CatBoost (AUC-ROC 0.844) and a CNN that estimates age from photos. Code and notebooks are in the TripleTenProgram repo.
Predicted contract cancellations so marketing could reach at-risk customers before they churned.
AUC-ROC ≥ 0.75
Random Forest · Logistic Regression · CNN · TensorFlow
Gradient-boosted model estimating used-car market value with a focus on prediction speed as well as accuracy.
Fast tuned GBM beat baseline linear
CatBoost · LightGBM · XGBoost · scikit-learn
NLP pipeline flagging negative movie reviews for sentiment tracking.
Multi-model NLP with transformer features
NLTK · spaCy · transformers · LightGBM
Hourly forecast of taxi demand to help optimize fleet scheduling.
RMSE ≤ 48 on hourly forecast
statsmodels · scikit-learn · LightGBM
Classification model that picks the best cell plan for a customer from usage patterns.
Accuracy ≥ 0.75
scikit-learn · Decision Tree · Random Forest · Logistic Regression
Predicted which bank customers were likely to leave so retention could be targeted.
F1 ≥ 0.59
scikit-learn · Random Forest · Logistic Regression
Evaluated borrower metrics to predict the likelihood of loan default.
Identified drivers of default risk
Python · pandas
Validated reserve volume models and calculated profit/risk trade-offs across candidate drilling regions.
Best risk-adjusted region selected
scikit-learn · Linear Regression · Bootstrap
Modeled the gold-mining recovery process to surface efficiency improvements.
Prototype model for industrial use
scikit-learn · Linear Regression · Decision Tree · Random Forest
Identified similar customers and predicted insurance benefit amounts while preserving data privacy.
Privacy-aware similarity + benefit model
scikit-learn · KNN · seaborn
Investigated which features drive used-vehicle prices in classified ads.
Feature-importance report for pricing
pandas · NumPy · Matplotlib
Tested hypotheses on user vs. critic scores to pick promising platforms and ad directions.
Statistically validated recommendations
pandas · SciPy · Matplotlib
Tested whether weather conditions meaningfully shift taxi trip duration.
Hypothesis test on real trip data
pandas · SciPy
Analyzed client behavior across telecom packages to identify which plans generate the most revenue.
Revenue-driver segmentation
pandas · SciPy · Matplotlib
Compared listening habits between two cities to surface patterns in music preference.
First end-to-end EDA project
Python · pandas
End-to-end DS workflow — cleaning, modeling, evaluation — focused on performance metrics and outcome prediction.
Full workflow, performance metrics reported
Python · scikit-learn · pandas