← All work

Project explorer

Every project, with a way to try it

Search by problem or technology. Each entry has a scenario to try, a command to run locally, where to look in the code, and what the project does not do.

These include web apps, local tools, research demonstrations and coursework, built with AI assistance. Run a command from a checkout of its repository after following that README; requirements differ, and demo database seed commands need a fresh path.

19 projects

Finance · Product & HCIWeb application

Frontier AI Valuation Lab

A financing workspace that connects capital objectives, ownership scenarios and a reviewable decision.

  • JavaScript
  • Node.js
  • Python
A scenario to explore
Compare hypothetical financing plans, change a retained-ownership objective and inspect where the ranking changes.
Follow the implementation
Follow the shared calculation engine and Node API, then compare with the independent Python reference and verification guide.
Project scope
Independent research with illustrative assumptions; company names imply no affiliation or endorsement.

Local review command

npm test
npm start
AI & search · Data & analytics · Backend systemsWeb application

LineageGuard Studio

Typed review workflows with evidence, policy checks and recovery.

  • Node.js
  • SQLite
  • JavaScript
A scenario to explore
Use the Studio Compare tab to choose two completed synthetic reviews, validate both exports locally and inspect their differences. Recorded decisions and current evidence freshness remain separate; no current run is required.
Follow the implementation
Follow one shared comparison report through the browser and file-only command. Review semantic inspection, bounded file reads, stale-read invalidation and offline checks that preserve existing run state.
Project scope
Synthetic demonstration; optional DataHub integration is read-only. Saved-export checks establish internal consistency, not source truth, authenticated approval or migration authority. Production behavior needs separate verification.

Local review command

node scripts/studio-review.mjs --scenario=safe-add --profile=standard --output=./.lineageguard/compare-standard.json
node scripts/studio-review.mjs --scenario=safe-add --profile=strict --output=./.lineageguard/compare-strict.json
node scripts/studio-compare.mjs --left=./.lineageguard/compare-standard.json --right=./.lineageguard/compare-strict.json
# Use new export paths when repeating; existing outputs are never overwritten.
Data & analyticsWeb application

SaaS Analytics Warehouse

A reproducible analytics pipeline for MRR, retention and unit economics.

  • dbt
  • DuckDB
  • SQL
  • Python
A scenario to explore
Build a fictional subscription warehouse, then trace one historical month from the MRR bridge to customer records, payment-failure losses and channel/reason totals.
Follow the implementation
Follow the read-only investigation through exact-cent customer-count, total-loss and payment-failure-subset reconciliation. Tests retain historical churn after reactivation and refuse broken joins or inconsistent source data before writing reports.
Project scope
Seeded fictional SaaS data; results describe the fixture, not a real business.

Local review command

uv sync --locked
make data build MONTHS=24 SCALE=0.25
uv run python -m saasreport.churn --month 2024-12 --output-dir ../tallyloom-churn-2024-12
Data & analytics · Backend systems · Product & HCIWeb application

Impact Ledger

Participation and spending reporting with a local API and atomic CSV imports.

  • TypeScript
  • Python
  • SQLite
A scenario to explore
Preview an attendance import, resolve validation errors, import it and compare unique participants with attendance and session spending.
Follow the implementation
Follow preview/import shared validation, exact money arithmetic and calendar-period comparison through API and browser tests.
Project scope
Hosted preview is read-only and synthetic. Writes require running the local application.

Local review command

python3 -m impact_ledger --db .local/demo.db seed
python3 -m impact_ledger --db .local/demo.db serve
Product & HCIWeb application

ContrastKit

Color contrast auditing across a web app, library and CLI.

  • TypeScript
  • React
  • OKLCH
A scenario to explore
Download a failing palette and selected pair as tokens JSON, audit it with the CLI, apply the suggested fix and confirm the next downloaded file passes.
Follow the implementation
Follow real browser JSON export into the built CLI. Tests preserve the pair and target, gate invalid names and keep actual colors unchanged by color-vision simulation.
Project scope
Color checks support accessibility work; they do not constitute a complete accessibility audit.

Local review command

npm ci
npm test
npm run build
Product & HCIWeb application

MOTIONX Studio

A motion-led studio website with a separate outreach-validation workflow.

  • Next.js
  • TypeScript
  • Python
A scenario to explore
Explore the responsive studio pages, then inspect how the separate Python workflow validates outreach inputs.
Follow the implementation
Follow the frontend structure and architecture notes; review validation boundaries independently of the public website.
Project scope
Public site and validation tooling are separate. A website preview does not demonstrate outreach delivery or client adoption.

Local review command

npm ci
npm run dev -- -p 3939
Backend systemsLocal API service

RelayBox

A durable signed-webhook inbox with idempotency and fenced worker leases.

  • Python
  • SQLite
  • HMAC
A scenario to explore
Deliver the same signed event twice, process the job and inspect retry, lease and recovery behavior locally.
Follow the implementation
Follow ingestion uniqueness, transactional lease ownership and worker receipt handling; inspect the local diagnostic dashboard.
Project scope
Local single-host SQLite service. Diagnostic samples are not throughput or service-level claims.

Local review command

python3 -m relaybox --db .local/demo.sqlite3 demo
python3 -m relaybox --db .local/demo.sqlite3 serve --demo-secret
AI & search · Backend systemsLibrary / CLI

TraceBench

Offline evaluation of recorded AI outputs with exact supplied-cost accounting and regression gates.

  • Python
  • JSONL
  • Decimal
A scenario to explore
Evaluate a bilingual recorded fixture, compare a baseline and inspect a deliberately failing regression or budget rule.
Follow the implementation
Read strict input contracts and the JSON/Markdown/HTML reports, then trace reusable CI gate behavior.
Project scope
Evaluates supplied recorded outputs; it does not call a model or measure live model quality.

Local review command

python3 -m tracebench evaluate --cases fixtures/cases.jsonl --outputs fixtures/clean.jsonl --pricing fixtures/pricing.json
FinanceLibrary / CLI

quantkit

Options pricing, portfolio optimization and backtesting checked against independent references.

  • Python
  • NumPy
  • SciPy
  • pandas
A scenario to explore
Generate fictional prices, estimate monthly risk-parity allocations using only prior closes and compare against equal weight on the same dates and costs. Inspect six CSV exports.
Follow the implementation
Follow estimation-window receipts, targets and wealth exports. Future-price perturbation tests keep earlier allocations fixed; the comparison includes shared warmup and entry costs.
Project scope
Educational quantitative software using illustrative inputs; no investment recommendation or live execution.

Local review command

uv sync
uv run python examples/walk_forward_risk_parity.py
Data & analyticsLibrary / CLI

powerlift

Experiment statistics with power analysis, CUPED, sequential monitoring and Bayesian tools.

  • Python
  • SciPy
  • statsmodels
A scenario to explore
Read a seeded per-user experiment CSV, check the allocation and outcome window, then compare raw and CUPED-adjusted estimates with uncertainty.
Follow the implementation
Follow the complete experiment readout from CSV validation through SRM and horizon gates to CUPED. Try the blocked scenarios before inspecting the statistical library references.
Project scope
Statistical results depend on study design and data assumptions; a significant result alone is not a business decision.

Local review command

uv sync
uv run python examples/experiment_readout.py --as-of 2026-09-15
Finance · Backend systemsLocal API service

ledgerline

A double-entry ledger API with immutable balanced entries and idempotent writes.

  • Python
  • FastAPI
  • SQLAlchemy
  • PostgreSQL
  • SQLite
A scenario to explore
Run a disposable local HTTP wallet: deposit, transfer, retry with the same key, reject an overdraft, reverse the transfer and reconcile current and historical balances.
Follow the implementation
Follow the one-command receipt: replay identity, unchanged rejected-write balances, reversal links, historical state and zero trial balance. The example owns and cleans up its temporary database and server.
Project scope
Engineering demonstration; it is not a bank, payment processor or accounting system certification.

Local review command

uv sync
uv run python examples/wallet_demo.py
AI & search · Backend systemsLibrary / CLI

searchlite

A full-text search engine with explainable ranking and retrieval evaluation.

  • TypeScript
  • BM25F
A scenario to explore
Compare two saved local indexes with your own relevance judgments in the browser, inspect query-level regressions and download the native-compatible report. Try the synthetic controls first, then use your own files.
Follow the implementation
The local-file worker shares validation, ranking and the strict per-query gate with the native command. In the real browser tests, a synthetic positive-average case still fails an individual regression. Follow report-byte parity, bounded inputs, cancellation and offline browser checks; separate execution and vector studies retain their original source revisions.
Project scope
Small in-memory library. Files stay in page memory; Clear drops application references without promising secure erasure. Exported reports contain query text and document IDs. Supplied labels/postings are not authenticated; synthetic judgments and timing fixtures do not establish general retrieval quality or production latency. Vector checks retain cancellation and rounding limits.

Local review command

npm ci --ignore-scripts
npm run build
searchlite_review_dir=$(mktemp -d)
searchlite_review_dir=$(cd "$searchlite_review_dir" && pwd -P)
node examples/search-docs.js index examples/fixtures/relevance-notes "$searchlite_review_dir/notes.json"
node examples/search-docs.js evaluate "$searchlite_review_dir/notes.json" examples/fixtures/relevance-notes/judgments.json
AI & search · Backend systemsLibrary / CLI

citecheck

Document question answering with retrieved citations, claim checks and abstention evaluation.

  • Python
  • Claude API
  • SQLite FTS5
A scenario to explore
Index the sample handbook, ask a supported question and compare with a question the documents cannot answer.
Follow the implementation
Follow real SQLite indexing, retrieval and sentence-citation checks. Markdown fence regressions verify that example headings remain under their enclosing source section.
Project scope
Offline fixture responses are hand-written. Live answers require provider configuration and may incur usage charges.

Local review command

uv sync
uv run citecheck index examples/handbook
uv run citecheck ask "What is the deadline for submitting an expense?"
AI & search · Product & HCIWeb application

ProofCast

Local media generation with provenance records and optional object storage.

  • Python
  • Streamlit
  • SHA-256
A scenario to explore
Generate a local asset bundle, verify the saved manifest against its actual media files, then alter a copied file and inspect the explicit failure.
Follow the implementation
Follow the offline verification CLI, byte digests and media-format checks. Reproduce missing-file, tamper and moved-bundle cases without uploading anything.
Project scope
Local workflow demonstration. A digest identifies bytes; it does not prove authorship or the truth of media content.

Local review command

.venv/bin/python -m proofcast.cli --title "Synthetic demo" --script "A local verification example."
Product & HCIDesktop application

Luma Keys

Offline piano practice with score views, planned fingering and graded sight reading.

  • Electron
  • React
  • three.js
  • Web Audio
A scenario to explore
Build a disposable browser session, generate an exercise, play clean and flawed timed runs, then inspect saved progress and repeat the exact seed.
Follow the implementation
Run five sight-reading stories and inspect screenshots, failure receipts, owned-server cleanup and Web Audio practice grading without optional engines.
Project scope
Browser QA uses synthetic exercises with 3-D hands off on macOS/Linux. Physical input, native packaging and optional engines need separate verification.

Local review command

npm ci --ignore-scripts
npx playwright install chromium --only-shell
npm test
npm run qa:browser -- --output-dir .qa-runs/reviewer-exercise
Product & HCI · Data & analyticsPlatform game

Blast Radius Daily

A one-minute Reddit puzzle about downstream effects of schema changes.

  • TypeScript
  • Phaser
  • Devvit
A scenario to explore
Pick a UTC incident day, place two circuit breakers, trace the real cascade, compare the computed result, move one synthetic vote, then delete and replay.
Follow the implementation
Open the playable offline build or produce its single HTML yourself. Follow the shared engine, exact source/build lock and maintained offline Chromium journey.
Project scope
Desktop/landscape pointer demo; fictional community data lives in tab memory and reload resets it. Comments are unavailable. Live Reddit/Redis behavior, keyboard-only canvas play and portrait layout require separate verification.

Local review command

npm ci --ignore-scripts
npm run build:demo
# Open dist-demo/blast-radius-demo.html from disk
npx playwright install chromium --only-shell
npm run test:demo
Finance · AI & searchResearch demonstration

Agentic Trading Research

An offline research architecture with a runnable fail-closed review gate.

  • Python
  • JSON Schema
A scenario to explore
Run six synthetic gate scenarios and inspect why missing or contradictory evidence blocks a research result.
Follow the implementation
Follow schema validation, gate reasons and negative tests; compare the architecture with the runnable subset.
Project scope
Research-review-only demonstration with synthetic fixtures; no automated orders, brokerage connection or trading advice.

Local review command

python3 -B examples/fail_closed_gate.py --all
python3 -B -m unittest discover -s examples -p "test_*.py" -v
Product & HCIWeb application

JUST-QR

A minimal browser QR generator with image export.

  • HTML
  • JavaScript
  • Canvas
A scenario to explore
Generate and export a QR code, try an invalid or oversized destination, recover with a valid URL and compare clipboard/download images.
Follow the implementation
Read URL validation and recovery in the single-page implementation, then inspect independent QR decoding and quiet-zone pixel checks for both export paths.
Project scope
Browser utility; a QR code encodes input and does not validate whether the destination is trustworthy.

Local review command

python3 -m http.server 8000
Product & HCICoursework

Teen Budgeting

A classroom budgeting application using fictional examples.

  • JavaScript
  • Node.js
  • HTML/CSS
A scenario to explore
Explore the budgeting flow and compare the displayed spending categories and totals with the source.
Follow the implementation
Read the course-project setup, server routes and frontend interactions; keep its classroom scope in view.
Project scope
Coursework demo with fictional data; no persistent authentication or real financial accounts.

Local review command

npm ci
npm test
npm run build
npm run dev