From Fourier analysis to video pipelines, a directory of tools you can explore and code you can inspect. Each project explains what it does, how it was checked, and where its claims end.
Public repositories
39
Browser experiences
28
Ways to explore
Math · software · data
39 public repositories · browse the complete collection below
A 15-factor equity risk model, brought out of the notebook and into an interactive portfolio workbench.
Python
TypeScript
Risk models
Personal research & implementation
Implementation, evidence & limits
The work
I built a factor model that separates those drivers, then made the saved model explorable without a Python environment. Edit positions, compare portfolios, and inspect the contribution of each factor to total variance.
Evidence
503 names, 1,937 observations, 252-day annualization, and the exact covariance settings.
Scope & limitations
This is a dated research snapshot, not a live risk feed or a return forecast. The universe uses current index constituents, raw-close returns have corporate-action limitations, and Size is a dollar-volume proxy. The demo uses the fitted base factors; it does not refit the model or run the repository’s separate forecasting experiments.
An interactive laboratory for Fourier decomposition: draw a waveform, reconstruct it with harmonics, and inspect amplitude, phase, and approximation error.
JavaScript
Fourier analysis
Numerical methods
Personal mathematical implementation
Implementation, evidence & limits
The work
I expanded an early Fourier visualization into a self-contained signal lab. It exposes the coefficients and connects harmonic selection to the reconstructed waveform and its error, so the mathematical explanation is tied to quantities you can inspect.
Evidence
Browser implementation, numerical tests, and the preserved original Fourier experiment.
Scope & limitations
This is a finite sampled signal model. Frequencies above the sampling limit can alias, drawn signals are discretized, and interpolation does not recover unobserved detail. Error metrics describe the current sampled reconstruction, not a continuous-time guarantee.
A browser extension that helps students explore co-op roles with transparent skill matching and optional on-device semantic scoring.
JavaScript
ONNX
Browser extensions
Personal product & implementation
Implementation, evidence & limits
The work
I built an extension that adds scoring and a full-page Job Explorer to NUWorks. It separates eligibility from relevance, shows matched and missing skills, and keeps the search usable while optional semantic inference runs in the background.
Evidence
Documents inference, calibration, privacy boundaries, and the Job Explorer workflow.
Scope & limitations
Match scores are relevance signals, not probabilities of getting hired. Semantic similarity can miss seniority, negation, and hard requirements. The production extension requires an authorized NUWorks session. Demo employers, roles, and résumé profiles are fictional; no live application is submitted.
A local browser workbench for inspecting firmware containers, reviewing byte changes, and exporting an auditable result.
Python
JavaScript
Binary formats
Browser workbench & tooling extensions
Implementation, evidence & limits
The work
I built a browser workbench around documented firmware codecs, with raw inspection, profile-gated resource editing, undo, and an export manifest. A synthetic example and before/after byte review make the editing workflow possible to explore without vendor firmware.
Evidence
Browser implementation, Python tools, tests, and documentation of supported operations.
Scope & limitations
A successful parse or export does not prove that a file is safe for a device. The example is synthetic and is not firmware. Unsupported mappings remain raw data, resized diffs are positional rather than semantic, and the workbench does not upload files or flash devices.
A reusable React library for Oura data, with a synthetic-data gallery and a more resilient loading pipeline for independent collections.
React
TypeScript
Data visualization
Personal library & API integration
Implementation, evidence & limits
The work
I built a React visualization library and expanded it with an inspectable synthetic gallery. The loader handles independent collection failures, pagination, and cancellation so charts can retain useful results without presenting stale data as current.
Evidence
React components, loading behavior, tests, and synthetic example.
Scope & limitations
The gallery demonstrates visualization and failure handling with synthetic data. It is not medical interpretation or a live view of an Oura account. Package consumers still need an authorized API integration; repository improvements may precede an npm release.
An interactive draft board that connects player projections, positional scarcity, roster fit, and a snake-draft schedule to an explainable decision.
Python
TypeScript
Decision models
Personal model & decision interface
Implementation, evidence & limits
The work
I turned a quantitative draft strategy into an interactive board. Visitors can mark players as drafted, track their roster, and inspect why the model values a candidate under the current draft scenario.
Evidence
Original Python strategy, projection snapshot, and interactive draft implementation.
Scope & limitations
The board reflects a ten-team, full-PPR league setup from September 1, 2026. Its 200 skill players support 14 quantified selections, followed by final-round D/ST and kicker guidance without player projections. It is not a live draft feed, a betting tool, or a guarantee of player availability; uncertainty proxies are illustrative assumptions.
A local contact-data workspace with CSV validation, personalized draft previews, duplicate checks, and controlled exports, built alongside an Electron enrichment app.
JavaScript
Electron
Data validation
Personal desktop app & data tooling
Implementation, evidence & limits
The work
I added a local contact studio that parses, validates, and previews the data before export. Its shared CSV parser also improves the desktop editor and merge flow, preserving quoted fields and embedded newlines through roundtrips.
Evidence
The browser workflow and dependency-free shared parser, validation, and template logic.
Scope & limitations
Email syntax checks do not establish deliverability or consent. The studio creates local draft files; it does not enrich contacts, create mailbox drafts, or send messages. The separate Electron application requires provider configuration for those integrations.
Inspect the decisions behind a classifier: thresholds, cohorts, errors, and the cost of a false positive.
Python
Evaluation
Data leakage
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
A reproducible Titanic experiment fits preprocessing only on training data. The browser explores saved held-out predictions, ROC curves, confusion matrices, and row-level errors.
Evidence
179 held-out passengers; 80.45% accuracy and 0.8394 ROC AUC on this split. Twelve Python and eight numerical JavaScript tests.
Scope & limitations
One historical passenger-level split, with no leaderboard or causal claim. Older MNIST and traffic experiments have separate documented boundaries.
Compose two videos, review a frame-exact cut list, and render vertical clips on your own computer.
Python
FFmpeg
Media pipelines
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
A local Python service handles bounded uploads, fit/crop previews, audio selection, cancellable FFmpeg jobs, and MP4 downloads. Every finished clip is checked with FFprobe.
Evidence
Fourteen Python tests and three browser workflows exercise actual rendering, frame counts, pixel layout, audio, cancellation, and downloads.
Scope & limitations
Requires local Python and FFmpeg. Ten-minute source limit; no URL downloader, captions, or social publishing. Original experiments remain in the history and inert archive.
Solve implied volatility, compare it with historical volatility, and inspect covered-call payoffs.
Numerical methods
Options
JavaScript
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
Black–Scholes inversion checks price bounds before solving. Editable close prices drive sample historical volatility; Greeks and payoff curves explain what changes with each assumption.
Evidence
Ten numerical tests, including inversion and payoff identities. Impossible premiums block analysis and export.
Scope & limitations
Synthetic scenarios and a European constant-volatility model. Payoffs are not backtests; American exercise, assignment, and live trading are outside the demo.
Explore six-channel sensor sequences and run the original saved neural-network weights in a browser worker.
TensorFlow.js
Web Workers
Signals
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
CSV validation, channel plots, exact 100-row windows, cancellable inference, threshold exploration, and JSON/CSV exports make the model inspectable.
Evidence
The actual 68,097-parameter model runs locally with artifact hash verification. Six contract tests and browser checks cover inference, imports, stale runs, and cancellation.
Scope & limitations
A research replay, with synthetic examples. Training sensor units and independent evaluation are unavailable; scores are not calibrated road-safety assurances.
Follow both sides of a hypothetical binary market through depth, size, costs, and capital carry.
Order books
Simulation
Python
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
Matched ask-depth fills replace headline-price arithmetic. Quantity sensitivity and a reconciled cost breakdown sit beside a searchable archive of 54 historical market descriptions.
Evidence
Ten JavaScript and six Python tests; thin liquidity and stale snapshots produce explicit failure states.
Scope & limitations
Order books are synthetic. Archived descriptions do not establish executable opportunities. Illustrative costs are not a current exchange fee schedule.
Compare frozen stock selections with consistent execution timing, costs, and attribution.
Experiment design
Portfolio math
Replay
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
A replay engine applies next-observation execution, fees and slippage, then exposes equity, drawdown, overlap, and position contributions. Experiments can be imported and exported.
Evidence
Eleven numerical and four Python tests. Identical selections reconcile to identical equity and 100% overlap.
Scope & limitations
The default selections and prices are hand-authored synthetic data, not GPT or Gemini output or evidence of model performance. Broker integrations are separate.
Turn a board image into a reviewable position, with confidence margins and keyboard correction.
OpenCV
Python
Human review
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
OpenCV template matching proposes independent square candidates. Crop and orientation controls preserve the source image; explicit FEN fields and python-chess validate the reviewed position.
Evidence
Fourteen Python tests. The synthetic starting board produces 32 pieces and 20 legal moves; missing-king and crop errors are surfaced.
Scope & limitations
Template-specific recognition, not an unseen-image accuracy benchmark. Position validity cannot prove the image was read correctly or recover move history.
Practice an answer, review it against a rubric, and keep a portable record of your revisions.
Practice tools
State management
Accessibility
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
Nine original prompts, topic filters, a timer, stable-ID drafts, review invalidation, and validated session import/export form an offline practice workflow.
Evidence
Eight JavaScript and five HTTP tests; browser verification covers persistence, edits after review, missing data, and mobile layout.
Scope & limitations
Rubrics are for self-review, not automated grading or hiring predictions. The separate legacy AI and speech integrations were not exercised with live credentials.
Explore a rack of letters, practice hidden solutions, and see which physical tile each answer consumes.
Tries
Web Workers
Algorithms
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
A trie search runs in a worker, while an independent Python counter solver checks parity. Repeated letters, blank tiles, hints, saved progress, and result exports are supported.
Evidence
Ten JavaScript and seven Python tests; the original 43,117-entry wordlist is retained. Three repeated letters map to distinct tiles.
Scope & limitations
The historical dictionary is not an official GamePigeon word list and has unverified provenance. The demo does not control a phone or mouse.
Build a focus routine with an inspectable timeline and a companion display for Mentra glasses.
TypeScript
Timers
Wearables
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
A shared deadline-based timer handles pause, resume, skipped rounds, and configurable breaks. The browser includes a real display preview and typed command simulator.
Evidence
Nine unit/host tests and four browser workflows, plus TypeScript, SDK manifest validation, and a portable Miniapp package.
Scope & limitations
Physical glasses and microphone were not tested. Browser reload resets a running session; the current Miniapp SDK is beta.
Turn saved public page content into source-linked contacts and individually reviewed email drafts.
Python
Parsing
Review workflows
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
Local extraction, URL validation, deduplication, editable templates, and stale-review checks lead to formula-safe CSV and individual unsent EML exports.
Evidence
Fifteen Python/HTTP tests and five browser workflows, including real draft downloads and narrow-screen layout.
Scope & limitations
A local Python service is required. Search is a manual handoff; the studio does not scrape live sites or send messages.
Explore how a browser extension can pause a request and preserve truthful cancellation behavior.
Browser extensions
HTTP
Accessibility
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
A bounded, accessible confirmation queue handles fetch and XHR, request aborts, timeouts, and unknown operations. The test bench uses the actual guard with an in-memory transport.
Evidence
Five classifier tests and nine browser workflows. Cancelling one queued request and approving another results in exactly one mock delivery.
Scope & limitations
The extension is a convenience layer, not a security boundary. No signed-in Instagram or installed-extension test was performed.
Trace a daily wellness summary back to the exact day and source coverage behind each statement.
API integration
Data quality
Python
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
Deterministic facts, explicit missing data, bounded pagination, token refresh, and delivery acknowledgments support a standalone day-by-day HTML explorer.
Evidence
Twenty-one Python tests and four browser workflows cover exact-day gaps, failures, exports, and responsive presentation.
Scope & limitations
The public demo contains fictional records. Authenticated Oura, AI, and Telegram services were not called during verification.
Compare saved catalog snapshots without confusing missing products, unknown availability, and restocks.
Parsing
Provenance
Python
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
A bounded non-executing extractor retains explicit availability evidence and decimal prices by currency. Standalone HTML reports support filters, paired snapshots, and evidence downloads.
Evidence
Twelve Python tests and four browser workflows. An unsupported current storefront is rejected instead of producing fabricated inventory.
Scope & limitations
The original storefront redirects as of October 16, 2026. Demo products are fictional; this is a saved-catalog workbench, not a live purchase bot.
Cent-level stake allocation, commission scenarios, complete-market checks, and a source-aware viewer repair the original arithmetic and make assumptions inspectable.
Evidence
Seven Python tests include Python/JavaScript parity. Equal-odds and three-way examples expose rounding and losing stakes.
Scope & limitations
Historical odds and synthetic examples only. Results assume matching settlement rules and accepted legs, with no wagering functionality.
Review comparable auction listings, their identity assumptions, and the effect of fees on a candidate.
Data validation
Comparables
Fork
Fork improvements · upstream csjh/SkyblockSniper
Implementation, evidence & limits
The work
Complete-snapshot validation and deterministic matching replace shared thread state. A peer audit preserves visible upgrades and lore, with searchable evidence and scenario costs.
Evidence
Eleven Python tests; the browser exposes peer comparisons, fee changes, empty filters, and export state.
Scope & limitations
Bundled auctions are synthetic. Asking prices are not sales evidence, and display identity cannot establish hidden NBT attributes.
Ask about a passage and keep selected responses in a searchable, opt-in notebook on the reader.
Lua
KOReader
Fork
Fork improvements · upstream drewbaumann/AskGPT
Implementation, evidence & limits
The work
Working masked provider settings, bounded requests, failed-history rollback, translation fixes, and namespaced modules are packaged with an explicit runtime allowlist.
Evidence
Sixty-eight LuaJIT assertions exercise actual plugin code with UI/network fixtures; all runtime Lua files compile.
Scope & limitations
No physical KOReader device or authenticated provider test. Keys and notes use the reader’s standard unencrypted LuaSettings storage.
Explore how pairwise choices change a ranking, with an offline fictional dataset for experimentation.
Elo
Transactions
Next.js
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
A default local lab makes the ranking mechanics visible. The opt-in database path uses signed vote tokens, replay protection, and atomic paired Elo updates.
Evidence
Seven model tests and five PostgreSQL integration tests include 32 concurrent votes, a same-nonce race, and forced-failure rollback.
Scope & limitations
Preference scores do not establish employer quality or individual outcomes. The new production migration was not applied during this update.
Compare housing-payment scenarios and inspect the inputs behind the monthly estimate.
React
Scenario modeling
Hackathon
HackUTD team: Dhyan, Ashwin, Spandan · scenario studio extension
Implementation, evidence & limits
The work
A local educational model replaces credentialed browser AI calls and approval claims. Rate, term, down-payment, and recurring-cost inputs feed saved comparison scenarios.
Evidence
Six model tests and two production browser workflows, including the zero-interest case and no external data requests.
Scope & limitations
An educational scenario tool, not a lending decision, quote, or eligibility assessment. No financial records or credentials are needed.
Explore the team’s original CAD models and sketch an educational autonomous route.
Three.js
Blockly
Team robotics
7110A team archive · original team attribution retained
Implementation, evidence & limits
The work
A filterable nine-model archive adds optional 3D loading, a Blockly/keyboard route editor, replay, validated import/export, and source-attributed code sketches.
Evidence
Six route-model tests and five production browser tests, including an actual archived GLB render and live deployment verification.
Scope & limitations
Historical team work. Large model downloads are optional, and generated routes are educational sketches without physical robot validation.
Replay motor commands and examine differential-drive geometry and heading-control assumptions.
Control systems
Kinematics
Team robotics
Archived team code, including KDalela · separate educational companion
Implementation, evidence & limits
The work
An offline companion integrates exact drive arcs, bounded heading control, editable geometry, and pose/CSV exports, with a concrete review of the original firmware behavior.
Evidence
Twelve math/control tests cover closed paths, settling, and timeout behavior. The companion’s source and hashes are pinned in the team website.
Scope & limitations
The original repository stays archived. The companion is published through 7110A-Website; its assumed plant is not hardware validation.
A maintained offline successor for studying retry behavior and bounded scheduling on a virtual clock.
Scheduling
Retry logic
Simulation
Fork history · kangoka/tiktodv3; new local educational successor
Implementation, evidence & limits
The work
The new reliability lab replaces the old engagement-service workflow with inspectable synthetic request traces, backoff, Retry-After, deadlines, and cancellation.
Evidence
Sixteen JavaScript and five Python tests cover Retry-After, deadlines, cancellation, and invalid input. Original archive bytes are verified against the historical commit.
Scope & limitations
Synthetic transport only. This does not contact a social service or reproduce the original automation; the fork’s upstream has no declared license.
An earlier portfolio preserved as a navigable record of projects and design evolution.
React
Archive
Information design
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
A searchable original-project directory adds filters, shareable views, project details, and reading-list export, with an explicit link to the current portfolio.
Evidence
Four model tests and four production browser workflows; the live Pages deployment is verified. Original sources and attribution remain visible.
Scope & limitations
Historical content is presented as an archive, not a current résumé or claim that old integrations are still operational.
A source-backed portfolio connecting project stories, working demos, and inspectable technical evidence.
Next.js
Provenance
Playwright
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
Curated case studies and this complete public directory share a responsive Next.js shell. Published demo bundles retain exact source commits and file hashes.
Evidence
Production browser checks cover routes, filtering, navigation, mobile layout, and the bytes of the published demo artifacts.
Scope & limitations
Demo snapshots are pinned releases; each project explains whether it uses synthetic examples, saved research, or a local runtime.
The public résumé source, with a reproducible publication path and a readable companion view.
LaTeX
Publishing
Reference
Personal project · implementation & maintenance
Implementation, evidence & limits
The work
The publishing work focuses on compilation, navigation, accessible reading, and artifact verification while preserving the existing biographical facts.
Evidence
The source and generated artifacts provide a reviewable record of the résumé rather than introducing new credentials or achievements.
Scope & limitations
Dates and incoming roles should be read exactly as stated in the source. This repository supports the portfolio rather than representing a separate product.
