Software with intent.
An umbrella for ambitious products. Where ideas turn into shipped software — and shipped software turns into impact.
Move fast. Validate aggressively. Build for impact.
Kodama Labs is a personal studio. Every product starts as a one-pager, gets a five-day prototype, and earns the right to keep living through paid traffic and real users.
We don't chase trends. We hunt for problems boring enough that nobody else wants to fix them, and important enough that the people who have them will pay attention.
The bar is honest software: tools that do what they claim, code that's a pleasure to maintain, and a process that treats time as the scarcest resource.
Ship first
A one-pager and a working prototype beat a perfect plan every time.
Validate hard
Paid traffic, real users, week-four go/no-go. No vanity metrics.
Own the stack
One monorepo, one engineer, one set of tools — radical simplicity.
Products
The ones that earned their place.
Unified dashboard for freelancers earning in multiple currencies. Tracks flows between business entities and personal accounts, categorizes statements with Claude AI, and visualizes cashflow with D3 Sankey diagrams.
Technical challenges
- Multi-currency accounting across business and personal entities, with manual rate overrides and timezone-aware periods.
- AI-assisted transaction categorization that learns from corrections without requiring a full ML pipeline.
- Real-time Sankey visualizations that survive arbitrary period clipping without distorting reserve semantics.
Impact
Replaces five spreadsheets with one source of truth for international contractors juggling business, personal, and investment flows.
Continuously ingests federal procurement data, flags unusual pricing patterns and political proximity, and produces neutral-language reports for journalists and citizens. Discipline: surface what's worth reviewing — never accuse.
Technical challenges
- Scaling Claude analysis across a continuous government dataset while staying inside Vercel's cron budget.
- Mapping political relationships (family, donations, wealth anomalies) without crossing into defamation territory.
- Multi-stage alert pipeline with deterministic deduplication across overlapping criteria.
Impact
Makes public procurement data legible to non-experts. Open infrastructure for accountability journalism.
Hypothesis → Validating → Validated → Shipped
Bets in flight. Each one runs the same playbook: a one-pager, a 5-day prototype, real ad spend, and an honest decision on week four.
Award flight concierge for Brazilian families.
Notifies families when award seats open on their preferred routes and dates across Azul, LATAM, and Smiles. Currently in validation: paid traffic running, manual concierge model, week-four go/no-go gates.
Next bet in scoping. Watch this space.
Next bet in scoping. Watch this space.
“If you can't articulate the riskiest assumption in one sentence, the idea isn't ready.”
Archive
A selection of older projects. Some shipped, some pivoted, all taught something.
Helped a Brazilian retail tech raise average ticket and cut attendance time by training an Apriori model that surfaces association rules to attendants in real time — used both as suggestions and as a smart cache for commonly bought-together items.
Healthcare MVP that processes ECG signals and classifies five arrhythmia categories in realtime. Built feature engineering on time-series ECG, wrangled PhysioBank into a labeled dataset, and trained a model that hit 85% accuracy across classes.
Email service that sends weekly clothing recommendations from multiple e-commerce sources. Web-scraped fashion catalogs, ran collaborative filtering on real users, and trained a churn model to keep the recommendations relevant over time.
Kaggle competition to predict how quickly pets get adopted, helping shelters improve profile appeal and reduce euthanization. Combined CNN+PCA image features, TF-IDF + TruncatedSVD on unstructured text, and structured-data feature engineering. Top 23% on public leaderboard.
Kaggle competition where the task was to help robots recognize the floor surface they're standing on using IMU sensor data. Heavy on time-series feature engineering, parameter tuning, and model evaluation. Top 100 on public leaderboard.
Let's build something worth shipping.
Open to interesting problems, collaborations, and conversations.