The Lab · Experiments & analysis

The work, run as experiments.

Every analysis here is framed the same way a scientist would: a hypothesis, the dataset, the method and why I chose it, the findings, and the takeaway. Code is linked. Where it’s live, you can run it.

Experiments

Hypothesis, method, result.

Three seeds below — placeholders mapped to my real domains. Each becomes a full write-up with live code and charts.

Forecastingscaffold

Forecasting peak-season outbound volume

Hypothesis: A seasonal time-series model predicts nightly outbound units more accurately than the planning team's rolling-average gut — early enough to staff against it.
Dataset: TODO — ~2 seasons of shift-level outbound (units, headcount, calendar features). Anonymized export.
Method: TODO — SARIMA vs. Prophet, backtested on held-out peak weeks. Chosen for interpretability and clean seasonality handling, so the floor can trust the number.
Findings: TODO — headline accuracy lift, e.g. “MAPE 18% → 7% vs. the rolling-average baseline.”
Takeaway: TODO — the operator payoff: staff to the forecast, not the gut, and stop over/under-planning peak nights.

preview

Actual vs. forecast — nightly outbound units

Python
pandas
statsmodels
Prophet

Run it · demo soonSource · repo soonData · data soon

Classificationscaffold

What actually predicts an inventory error?

Hypothesis: A handful of features — location type, SKU velocity, shift — explain most reconciliation errors, so we can target the fix instead of auditing everything.
Dataset: TODO — labeled reconciliation events across a portfolio period. Anonymized.
Method: TODO — logistic regression + SHAP for ranked, explainable drivers. Picked because ops stakeholders need to *see why*, not just a black-box score.
Findings: TODO — top drivers + recovered value, e.g. “3 features → 80% of error variance; $X recovered.”
Takeaway: TODO — where to point the reconciliation process first for the biggest accuracy gain per hour.

preview

Ranked feature importance — drivers of error

Python
scikit-learn
SHAP
SQL

Run it · demo soonSource · repo soonData · data soon

Causalscaffold

Does the trade journal actually improve results?

Hypothesis: Journaled, rule-checked trades beat impulse trades on win-rate — the discipline, not the setup, is the edge.
Dataset: TODO — personal trade log from Automate Ascension (journaled vs. un-journaled cohorts).
Method: TODO — cohort comparison with bootstrapped confidence intervals. Chosen to put an honest error bar on a small, self-collected sample.
Findings: TODO — win-rate delta + CI, e.g. “+11pp win-rate, 95% CI [4, 18].”
Takeaway: TODO — turns a trading belief into a measured behavior — the analyst instinct applied to my own decisions.

preview

Win-rate by cohort — journaled vs. impulse

Python
pandas
NumPy
matplotlib

Run it · demo soonSource · repo soonData · data soon

Iterations

The work in motion.

Screens and sketches from build to shipped — proof the polish was earned, not assumed.

screenshot

v1 — sketch

screenshot

v2 — wired

screenshot

Shipped

Source

On GitHub.

The repos behind the work. Demo where it's deployed, source where it's public.

All repos

dysick-portfolio

This site — Next.js 16, Tailwind v4, a hand-built design system.

Next.js
Tailwind v4
TypeScript

Demo · soonSource · soon

compass

Personal-finance command center — Plaid sync, debt engine, FIRE projection.

Next.js
Supabase
Plaid

Demo · soonSource · soon

automate-ascension

Self-built trading desk — live charts, scanners, disciplined journal.

React
FastAPI
Postgres

Demo · soonSource · soon

Get in touch

Let’s find the next best move.

Hiring for an operations, analytics, or data role — or have something that needs building? Either way, the fastest path is a direct email. I read every one.

Based in: Atlanta, GA
Working: Remote-first
Open to: Senior ops / analytics / data
Studio: DVSN × ZERO