Case studies

Systems that run without me.

Each project below runs in production, unattended. Every case surfaces a live demo and the code within the first screen — problem, architecture, decisions, metrics.

FIELD NOTE // 01Human direction. Autonomous execution.Designing the interfaces, guardrails, and feedback loops that move agent systems beyond the demo.
Agent tooling · GitHub automation · instruction synthesisSkill MergeScan real Hermes skill Markdown files in a GitHub repo, detect section conflicts, resolve them side-by-side, and open a draft PR with one unified AGENTS.md.25 TDD tests · real draft PR outcomeBrowser media engineeringKeeptakeMark useful moments in a real local screen recording, then render one playable visual WebM cut with bounded context and no upload.5 unit and language tests · 15 browser testsDesign-system accessibilityLuma WeaveA browser-local contrast cartographer that maps every CSS color-token pairing and writes a minimal AA repair back into source.4 TDD · 7 browser stepsLocal voice authoringWallnotePair one live public-domain museum work with a real microphone take, then share or download the source-linked voicecard without uploading voice bytes.7 unit · 15 browser checksLocal image geometrySketchshiftTurn a real photo of a paper or whiteboard flow into editable, accessible browser geometry without uploading the image.5 unit · 25 browser checksIncident process studioHandrailTurn an incident lead's own response steps into a persistent, keyboard-ready process rail that can be shared without an account or backend.12 unit · 8 browser testsRepository cartographyChorus MapTurn one live public GitHub repository into a portable architecture poster without cloning or executing its code.5 unit · 2 browser testsGitHub transaction safetyDecision PressTurn one merged pull request into a source-linked architecture decision, then open it as a draft review.14 tests · 5 browser modesProbabilistic planningWeather Window BookerTurn one outdoor task and real place into a calendar-ready work block backed by a live WeatherNext ensemble.27 unit · 8 browser testsGitHub transaction safetyRedline ReproTurn one issue into one bounded regression-test draft pull request without executing or merging repository code.21 unit · 8 browser testsCI recovery systemsNight RequeueReview one exact GitHub Actions attempt, isolate every non-success job, and requeue failures only after the live run identity and affected job set still match.1 real production requeueMarket data systemsSeamlineNormalize public Polymarket and Kalshi books, reconstruct a three-source fair line, and fail closed before recording a bounded browser-local paper intent.33 unit tests passingReal-time systemsSidebandConnect two browsers with human-carried WebRTC signaling, then pass live text and files directly without accounts or payload storage.10 unit tests passingRelease automationRelease PagekeeperTurn a live GitHub release and one stale Markdown block into a deterministic branch, exact-SHA commit, and reviewable pull request.25 unit tests passingLocal-first file operationsDownloads CrewStage loose local files into predictable folders, verify every copy, and remove originals only after the whole batch is safe.18 unit tests passingStorage reliabilityParity DeckCycle six erasure-coded shards through healthy, missing, and corrupt states, then reconstruct from any four checksum-valid survivors.7 engine tests passingLocal-first planningRetainer RunwaySplit a prioritized client worklist into what fits a fixed monthly retainer and what overflows.4 allocation tests passingRelease automationBumpdockTurn a live monorepo pull request into one dependency-aware Changesets commit guarded by the exact reviewed branch head.31 automated checks passingLocal-first browser adaptationTapReadyEnlarge one hard-to-hit live-page control through a reversible, origin-local bookmarklet without uploading page data.24 automated checks passingBrowser document engineeringFinal PacketTurn real local PDFs and images into one indexed, submission-ready packet without uploading source bytes.26 automated tests passingVoice AI productAiVoicePhoneA 24/7 AI receptionist that answers the calls a business would otherwise miss and turns them into captured, qualified leads.24/7 live answeringAI agencyTreCal Media GroupAn AI agency delivering automation, lead generation, and customer-communication systems to realtors and small business owners.AI systems, deployedBrowser media engineeringClean TakeDraw time-bounded privacy tape over a real local video, optionally mute the interval, and render a fresh WebM without uploading source bytes.13 browser tests passingLocal-first developer toolingRuleboundBind an explicit engineering policy to a real local JavaScript repository, preview current breaches, and install a runnable gate without uploading files.2 production browser paths passingModel evaluationAI Code JudgeCompare code-generation systems with executable rubrics, unbiased pass@k, paired bootstrap intervals, and reproducible evidence receipts.Paired bootstrap + pass@kAgent coordinationSecond ShiftTurn a live public GitHub issue queue into one collision-resistant work floor shared across real browser tabs.10 core tests passingPrivacy engineeringNoise LedgerCompose repeated Gaussian analytics in Renyi space, inspect each query's exposure, and gate the full plan before release.6 release tests passingSoftware supply-chain securityAttest LensA browser-native provenance microscope that separates signed-byte integrity, artifact binding, policy, and fail-closed promotion.6 release tests passingDistributed systemsLease ForgeConfigure a lease expiry, replay one resumed stale command, and inspect why the resource must fence old holders.Token 1 rejected after token 2Prediction-market intelligenceSmart Money TrackerA read-only decision surface that normalizes current Kalshi and Polymarket activity into one tape for whale prints, leaderboard identities, wallet flow, and smart-money agreement.up to 400 trades / snapshotOrchestrationAgentic WorkforceA personal team of specialist agents run by a chief-of-staff, reachable from my phone.11+ agents liveAutonomous pipelinesOvernight FoundryAgents research, spec, and advance business ideas overnight; a brief is on my phone by 8am.3 ideas/wk vettedReliabilitySelf-Improvement LoopA daily loop where agents propose upgrades and safely patch their own prompts with rollback.snapshot + rollbackContent automationSEO / Content SystemsAn automated content pipeline plus continuous rank and keyword monitoring on top.pipeline → Canva → calendarLocalization toolingPlural LensA cardinal-plural proofing workbench that makes missing locale forms visible across 201 integer counts.6 locales · 201 countsDeveloper toolingCache WindowA response-cache policy instrument that turns Cache-Control directives into a scrub-able freshness and stale-reuse timeline.10 core tests passingSecurity toolingCell SentryA local CSV inspection workbench that maps formula-capable prefixes to exact parsed cell coordinates.11 core tests passingDistributed-team toolingOverlap LoomA recurring-meeting loom that exposes daylight-saving offset changes across eight weekly local-time outcomes.9 core tests passingReliability engineeringRetry YardA deterministic webhook delivery simulator that makes backoff, throttling, idempotency, and dead-letter routing inspectable.16 behavior tests passingDistributed systemsCausal CanvasA local-first synchronization microscope that makes causality, concurrency, and deterministic convergence visible.7 automated tests passingObservability engineeringTail ScopeA deterministic tail-sampling policy lab that makes ordered decisions and bounded capacity pressure inspectable.6 behavior tests passingComputational imagingImageManipA serverless Python pipeline for controlled frequency, sensor-noise, compression, entropy, and metadata transformations.4 camera profilesNLP · YouTube API · Next.jsClapmarkTextTiling NLP segments a real ASR transcript into topically coherent chapters. TF-IDF names each segment. One button writes the chapter block to your YouTube video description via OAuth.8 TDD behaviors passingSecurity · Binary Analysis · Next.jsMIME DispatchDrop any file and instantly compare its magic-byte signature, file extension mapping, and browser-reported MIME type. Detects file-type masquerade attacks — ELF binaries disguised as images, polyglot ZIPs, misnamed media — with a hex dump showing exactly which bytes triggered the signature match.11 TDD behaviors passingHCI · Ergonomics · ISO 9241-411 · Next.jsfitts-labRun a live ISO 9241-411 pointing calibration session at two DPI conditions. The lab measures your actual movement time per trial, applies least-squares Fitts's Law regression, filters IQR outliers, and recommends the throughput-optimal DPI with an OpenLogi-compatible profile export.13 TDD behaviors passingGeospatial · Browser-local · SVG · Next.jsGeoStampLoad your Google Timeline export, and GeoStamp detects distinct trips using a home-cluster spatial algorithm, renders each trip on an animated equirectangular SVG map, and exports a shareable travel card — all without uploading a single byte.23 TDD behaviors passingAPI tooling · developer productivitySchema LedgerPaste two JSON Schemas and get an exact backward-compatibility breakdown — removed fields, type changes, narrowings, new required constraints — each with a TypeScript migration guard stub. Browser-local, zero API calls.12 TDD unit tests · 0 external API callsdeveloper tooling · CI/CD · graph algorithmsFlow FormEnter any public GitHub repo and get a topological CI dependency diagram committed directly to /docs/flow-form.svg — no Mermaid, no graphviz, custom DAG layout engine from scratch.17 TDD unit tests · custom Kahn topological sort + SVG layoutDeveloper ToolCron PlotPaste any 5-field cron expression — see its 90-day week × hour heatmap and detect midnight storms, weekend gaps, and rate spikes instantly.21 TDD testsDeveloper tooling · log analysis · browser-localSignal PlotPaste nginx or Apache access log lines. Get instant operational intelligence: request rate timeline, status breakdown, top paths, and anomaly detection — all browser-local.26 TDD tests · browser-local · 0/10 slopAI tooling · image analysis · browser-localChroma StampDrop any local image and get its visual DNA — palette, edge character, tonal key, and composition — as a ready-to-use GPT-Image-2 style prompt. Entirely browser-local, zero upload.26 TDD tests · median cut quantization · browser-localDeveloper ToolQuery SlicePaste any multi-table SQL SELECT query and get an instant JOIN graph, predicate inventory, output column map, and complexity score — browser-local, zero upload.31 TDD testsDeveloper Tool · Graph Algorithms · GitHub APIMesh GuardFetch real import graphs from any public GitHub repo, detect circular dependency cycles with Tarjan SCC, see Martin instability per file, draw boundary zones, and export an enforceable ESLint config.34 TDD tests · Tarjan SCCDeveloper ToolLog LedgerPaste a conventional-commit git log and get instant semver bump detection, commits grouped by type, and a ready-to-copy CHANGELOG.md block — browser-local, zero uploads.25 TDD testsKnowledge EngineeringWikiWeaveLoad your local Markdown vault, compute TF-IDF cosine similarity between every note pair, and write [[wikilinks]] directly into your source files — no upload, no server, no cloud.31 TDD testsSpace Tech · Orbital Mechanics · Satellite TrackerOverpassEnter your location, pick a prediction window, and Overpass computes every upcoming satellite pass overhead — using real live TLE data and SGP4 propagation running entirely in your browser.20 TDD tests · SGP4 · ICS exportDeveloper Tool · WebAssembly · Binary InspectorWasm ScopeDrop any .wasm binary. Decode its section manifest, type signatures, imports, and exports entirely in your browser — no upload, no execution, no server.39 TDD tests · zero executionDeveloper Tool · HTTP · Security AnalysisHTTP TracePaste any raw HTTP/1.1 request and get instant anatomy: headers annotated with RFC semantics, body decoded (JSON/form), and an A–F security scorecard — entirely in your browser.30 TDD tests · Command/Inspect surfaceDeveloper Tool · Text Analysis · LCS AlgorithmDiff MotionPaste any two texts and watch character-level changes appear instantly: deletions in red, insertions in green, with edit distance, LCS length, and similarity percentage — computed entirely in your browser.24 TDD tests · LCS-optimal diffMonitor · Data Visualization · Real-Time APIQuake PulseLive USGS M2.5+ earthquake feed plotted on a rotating orthographic SVG globe with depth classification, magnitude-coded markers, event sidebar, and time/magnitude filters — computed entirely in the browser, zero upload.32 TDD tests · Live USGS feed · SVG globeDeveloper Tool · Observability · ArchitectureArc TraceUpload an OpenTelemetry trace export and commit a runtime service topology map to your GitHub repository — your actual architecture, derived from production telemetry, not a hand-drawn guess.27 TDD tests · Operate surfaceDeveloper Tool · Typography · Binary ParsingFont ProbeDrop any TTF or OTF. Font Probe decodes every SFNT binary table in pure TypeScript, surfaces Unicode block coverage across 52 blocks, classifies the license type from name table text, and renders your text live in the loaded font — zero upload, zero server.20 TDD tests · Command/Inspect surfaceDeveloper Tool · GitHub API · Import Graph AnalysisBranch DeltaPaste a GitHub PR URL. Branch Delta fetches both commit trees via the GitHub Trees API, extracts import statements from every changed TypeScript, JavaScript, and Python source file, computes the structural dependency delta, and posts a formatted review comment directly to the PR.42 TDD tests · Operate surfaceSecurity Tool · Browser-local · Content Security PolicyCSP ProbePaste any Content-Security-Policy header and get instant forensic anatomy: directive inventory with per-source risk classification, attack-vector coverage matrix, A–F security grade with scored deductions, and actionable hardening steps. Browser-local. Zero upload.44 TDD tests · Command/Inspect surfaceGitHub Discussions · GraphQL API · Community toolingStudy SyncFetch the most-engaged Discussion threads from any public GitHub repository, compose a structured analysis/question/answer in a two-panel workspace, and post a real GitHub Discussion comment via the GraphQL addDiscussionComment mutation.40 TDD tests · real Discussion comment publishedWebMIDI · Music Theory · Real-time Signal ProcessingNote RollPlay notes on an on-screen piano (or MIDI keyboard). Note Roll captures the live performance, runs Krumhansl-Schmuckler key detection against 24 psychoacoustic key profiles, segments chord clusters, and exports a downloadable SVG lead sheet — all browser-local, zero upload.34 TDD tests · Krumhansl-Schmuckler 1982 algorithmFacilitation Tool · Browser-local · Timebox PlanningAgenda LoomShape a meeting around one fixed hard stop, adjust each segment in place, and expose breathing room, overruns, and after-stop work before the session begins.Live overrun classification · zero calendar writesImage Tool · Browser-local · Responsive Art DirectionCrop ChoirPlace one focal point on a local image, compare it across four publishing frames, and hand one portable CSS object-position recipe to a site or CMS.4 responsive apertures · 15 MB local input capDeveloper Tool · Protocol Buffers · Browser-localProto ScopePaste any base64-encoded protocol buffer binary and an optional .proto schema. Instantly decode every field: wire type, byte offset, raw value, and schema-matched name with type interpretation — all in the browser, zero upload.27 TDD tests · pure TS varint/wire decoderHealth Tool · Signal Processing · Browser-localCadence LogOpen on your phone and walk or run. The browser reads real accelerometer data at 60–120 Hz, applies a 4th-order Butterworth filter and an adaptive peak detector, counts each step, and downloads a GPX 1.1 cadence log you can import to Strava, Garmin Connect, or any fitness app.38 TDD tests · 4th-order Butterworth filter + adaptive peak detectorDeveloper Tool · RFC 7519 · Browser-localJWT ScopePaste any JWT token and get instant forensic anatomy: algorithm risk classification (none=CRITICAL, RS256=STRONG), RFC 7519 registered claim annotations with semantic descriptions, live expiry countdown, OIDC extension claims, and key ID detection — all in the browser, zero upload.37 TDD tests · pure TS JWT parserDeveloper Tool · npm Registry · Compare SurfacePkg DeltaType a package name and two version numbers. Pkg Delta fetches both manifests live from the npm registry and renders a structured diff: dependency additions, removals, semver-classified bumps (MAJOR/MINOR/PATCH), script changes, and manifest field changes — entirely in the browser, zero backend.35 TDD tests · pure TS diff engineAudio Tool · Web Audio API · Operate SurfacePod ChapterDrop a podcast or audio file. Pod Chapter decodes it in your browser using the Web Audio API, runs an energy-envelope silence-detection algorithm to find natural topic breaks, shows an editable chapter timeline, and exports an MP3 with embedded ID3v2 CHAP frames — no upload, no account, no server.32 TDD tests · ID3v2 CHAP encoding · zero uploadCode Review Tool · Static Analysis · Operate SurfaceYAGNI WatchPaste a GitHub PR URL. YAGNI Watch fetches the unified diff, runs static analysis on TypeScript, JavaScript, and Python files to detect YAGNI violations — dead conditionals, TODO stubs, over-parameterized functions, single-variant switches, unnecessary I-prefix interface abstractions — and posts a structured GitHub Review with file:line citations.24 TDD tests · 5 violation categories · GitHub Review postPerformance Tool · Network Inspector · Command/Inspect SurfaceHAR InspectDrop a browser HAR file and instantly see an interactive waterfall timeline of every request — timing phases (DNS / TCP / TLS / TTFB / receive), response sizes, speed classification (FAST/OK/SLOW/CRITICAL), domain grouping, and the three slowest requests highlighted — all browser-local, zero upload.25 TDD tests · 7 timing phases · zero uploadWriting Tool · AI Pattern Detector · Operate SurfaceProse SharpPaste or type prose. Prose Sharp detects 12 specific AI-writing pattern types inline in real-time — hedging, filler intensifiers, passive constructions, pseudo-profound qualifiers, and more — highlights each match with colored wavy underlines, and lets you accept or reject targeted fixes one annotation at a time. A paragraph-density heatmap rail shows pattern concentration at a glance. Browser-local, zero upload.22 TDD tests · 12 pattern types · browser-localPrint Utility · Imposition Tool · Operate SurfaceFolio FoldBuild a duplex booklet imposition plan in one browser tab. Name each page, set reading direction, and Folio Fold computes the physical sheet order for saddle-stitch binding — showing exactly which pages share each face, generating a print-room ticket, and exporting a JSON plan. No upload, no document parser, no account.browser-local · saddle-stitch imposition · JSON exportVideo Tool · Caption Tool · Operate SurfaceBreathlineTurn one spoken passage into timed, readable caption cues and export a browser-local WebVTT file. Tap any word to set a cue break; the timing strip updates instantly to show duration and words-per-minute for every cue — no upload, no backend, no account.browser-local · WebVTT export · wpm analysis