Most tools tell you what happened. Mine tell you why — and admit it when they're guessing.

[tkwind]
now      = "building developer tools that explain themselves"
ask_me   = ["CLI design", "auditable LLM agents", "why your API returns 405"]
weekends = "tuning a 3D printer — and the slicer fork it demanded"

⚑ Flagship — repoclean · pip install repoclean-cli — a pre-commit gatekeeper that pairs regex token detection with entropy scoring, blocks the leak in strict mode, and emits JSON for CI.

• PostSense — diffs a failing request against your last working one, and simulates the browser CORS constraints your desktop client hides; every diagnosis carries an evidence grade.
• Apply_AI — React Query + JWT auth over an Express/Mongo backend, with NVIDIA NIM parsing job descriptions into resume suggestions.
• ai-outreach-assistant — scoring, messaging, and orchestration kept deliberately separate so every decision is auditable; runs on local Mistral, so lead data never leaves the machine.

repoclean  ·  A git hygiene scanner that installs itself as a pre-commit gatekeeper. It catches GitHub, Slack, Stripe, Telegram, AWS, and OpenAI tokens plus high-entropy assignments, and strict mode blocks the commit outright. JSON output plugs into CI. The premise: most leaks are speed, not carelessness — so hygiene has to be automatic.

PostSense  ·  An API client that debugs instead of reporting. It compares a failing request to your last successful one for the same endpoint and returns a diff, not a status code. It also auto-probes unknown endpoints in a rate-limit-safe sequence and simulates browser CORS constraints. Every diagnosis is graded High / Medium / Low, so you always know inference from knowledge. Single folder, vanilla JS, no install.

Apply_AI  ·  A job application tracker with a drag-and-drop board. NVIDIA NIM parses job descriptions into resume suggestions. React + TS + Vite on the front, Express + Mongo behind JWT auth, React Query holding the two together.

ai-outreach-assistant  ·  Agentic B2B lead scoring built as three deliberately separate pieces: a scoring agent, a messaging agent, and an orchestrator. Decisions stay apart from execution, thresholds stay configurable, and every call leaves an explainable trace. It runs entirely on local Mistral via Ollama — no lead data leaves the machine.

Not everything in an account is a product:

• jatayu-fastapi-crud — FastAPI + Firestore task API, Dockerized for Cloud Run, with Firestore access isolated in a service layer and a folder structure deliberately kept readable for people learning the stack.
• pharmazephyr-backend — Node service, in progress.
• OrcaSlicer-bambulab — the slicer fork my printer demanded.

The same decision loop ships in all three tools:

%%{init: {'theme':'base','themeVariables':{'primaryColor':'#12293B','primaryTextColor':'#7FD8D2','primaryBorderColor':'#7FD8D2','lineColor':'#FFB86B','secondaryColor':'#0D1B2A','tertiaryColor':'#0D1B2A','fontFamily':'ui-monospace, monospace'}}}%%
flowchart LR
    A([failure signal]) --> B{known-good<br/>baseline?}
    B -->|yes| C[differential analysis]
    B -->|no| D[heuristic inference]
    C --> E[HIGH<br/>observed evidence]
    D --> F[MEDIUM / LOW<br/>flagged as guesswork]
    E --> G([one-click fix])
    F --> G
    G -.-> H[becomes the next baseline]
    H -.-> B

differential analysis → PostSense's request diff  ·  heuristic inference → repoclean's entropy scoring  ·  the explainable trace → ai-outreach-assistant's audit trail

rendered in-repo by scripts/render_cards.py from live GitHub API data — no third-party stat services anywhere on this page

No stat-card services, no rate limits, nothing that breaks at busy hours.

• The header (ASCII sweep + typing tagline), terminal, stack grid, badges, and divider are hand-written animated SVGs in assets/ — CSS keyframes inside the files. GitHub renders them through its image proxy, so the animations run but scripts never do.
• The stat dashboard and repo cards are rendered by scripts/render_cards.py from live GitHub API data.

Theme switching uses <picture> + prefers-color-scheme, so every graphic ships in a dark and a light version.