Lalabase
티켓·문서·근무시간·고객대화를 한 곳에 모아 KI로 검색, 소프트웨어 기업 고객 프로젝트용, 오픈소스로 자체 서버·독일 클라우드에 배포
개인 계정의 공개 저장소 소유자는 GitHub로 로그인해 운영자를 확인하고 소개를 직접 관리할 수 있습니다. 관리자 승인은 필요하지 않습니다.
로그인한 뒤 확인 파일을 만들어 이 저장소의 기본 브랜치에 30분 안에 커밋해 주세요. 다른 브랜치나 포크의 파일은 인정하지 않습니다. 파일을 올린 뒤 아래에서 확인하면 관리자 승인 없이 관리할 수 있습니다. 자격을 유지하려면 파일을 남겨 두세요.
로그인이 필요하면 로그인 화면으로 이동합니다. 로그인 후 이곳으로 돌아와 원하는 확인 버튼을 다시 눌러 주세요.
프로젝트 폴더에서 /nomorevibe verify를 실행하면 소유권을 확인하고 소개를 직접 관리할 수 있습니다.
1분 만에 확인하기Forecast-informed scheduling for Miami outdoor crews, powered by Gemma 4.
저장소 README에서 가져옴
README 전체를 불러오는 중입니다.
README 전문 보기 ↗A clearer workday for Miami's outdoor crews. Kruu turns a work order into a forecast-informed schedule, helping supervisors move flexible outdoor tasks away from the hottest hours while preserving the work, fixed commitments, breaks and paid shift.
Run locally · How the AI works · Verification · Submission
The work stays. The timing can change. Kruu puts the original and proposed day on the same timeline, with appointments, breaks and paid hours preserved.
Captured from the production app on October 10, 2026, using a fictional work order and the live forecast available at capture time. The illustrated 150 → 0 comparison measures outdoor minutes in the three hottest forecast hours; it is one example, not a guaranteed result or a health outcome.
Watch the story: Faye’s final narrated demo, with project links and contributor credits. Release page and download.
Calendar exports include the full paid shift, proposed tasks, fixed break, paid availability and selected-language briefing. Times follow America/New_York, including daylight saving. Exports are snapshots; later changes do not sync automatically.
The prototype supports one crew, one Miami forecast location, one date and up to four tasks. The supervisor makes the final decision. Kruu supports planning; it does not certify that a shift is safe or claim a measured reduction in injuries.
The work order sits beside its forecast context. Gemma extracts tasks for the supervisor to review.
Switch between English and Español, review the briefing, then copy it, download it or take the plan to your calendar.
Production captures from October 10, 2026. Forecasts and generated wording reflect that recorded example, not current conditions.
The Miami heat guide gives residents a direct way to check today's available forecast and find county-listed places to cool down—without a work order or account.
The server validates the county table and reuses successful checks for up to five minutes within a warm server process. If the source is unavailable or its structure changes, Kruu links directly to the official directory. Listings do not confirm current opening, accessibility or capacity. The guide does not assess personal risk or provide active alerts.
Kruu follows the Soft Signal design direction: warm ivory and cream surfaces, apricot controls, cocoa typography and a coral-to-orange ripple sun. The lowercase wordmark and generous spacing keep the interface approachable, while source timestamps, task constraints and uncertainty stay visible.
The application mark is a documented geometric reconstruction of the approved mockup, paired with a Manrope wordmark. The forecast gradient illustrates apparent temperature; its colors are not safety thresholds.
Runtime model: gemma-4-26b-a4b-it, through the Gemini API. Kruu calls the model server-side using Google's official @google/genai SDK. There is no model substitution or canned-response fallback.
| Responsibility | Implementation |
|---|---|
| Extract editable tasks and flag missing details | Gemma 4, with schema validation and numbered work-order citations |
| Check feasibility and compare schedules | Deterministic TypeScript |
| Explain the computed result and prepare bilingual briefings | Gemma 4, using the confirmed schedule |
| Protect the comparison passed to the briefing endpoint | Server-signed results; altered or expired data is rejected |
The server copies cited source text from the original work order for review. Missing inputs and unsupported timing precision require supervisor correction. Gemma does not decide schedule feasibility or calculate the comparison metric.
Explore the model integration, schemas and scheduler. Model documentation: Gemma through the Gemini API and Gemma 4 model card and license.
The primary metric is outdoor minutes overlapping the three hottest hourly apparent-temperature periods that intersect the paid shift. Tied temperatures select earlier hours consistently; partial hours count only the minutes inside the shift. This compares forecast exposure, not personal heat risk.
Every task is a contiguous block on a 30-minute grid. Tasks cannot overlap each other or the fixed break. The scheduler preserves durations, time windows, fixed appointments and dependencies. Ties favor lower duration-weighted apparent temperature during outdoor work, then fewer changed tasks, then less total movement. A change that improves only average temperature is labeled an optional alternative.
A valid original schedule is required for an honest comparison. Missing details, unsupported precision, conflicting appointments or incomplete weather coverage block the result. Unchanged and infeasible plans are valid outcomes. Open shift time remains paid availability.
Kruu uses the National Weather Service API, with Miami grid MFL/110,50, derived from 25.7617, −80.1918:
The adapter converts explicit Celsius/Fahrenheit units, maps timestamps to America/New_York, checks complete hourly coverage and displays the older source timestamp. Data older than 12 hours is refused. NWS responses may be reused for up to five minutes. Repeated or skipped local hours at daylight-saving transitions are refused rather than guessed.
Forecasts do not represent every job site's shade, exertion, equipment, humidity exposure or worker condition. Kruu does not replace workplace procedures or the supervisor's judgment. No worker interviews or health, productivity or financial outcomes have been measured.
Work-order text is saved locally in the browser and sent to the configured model service for extraction. Avoid worker identities and private client information. Kruu includes no accounts, database, analytics or automatic crew messaging. Model request sizes and durations are bounded. Rate limiting is best-effort per instance, not a distributed abuse or billing control.
Use Node.js 24 LTS (recommended), npm and a Gemini API key with access to gemma-4-26b-a4b-it.
git clone https://github.com/NaxeCode/heatshift.git kruu
cd kruu
npm ci
cp .env.example .env.local
# Add your own GEMINI_API_KEY to .env.local in an editor.
chmod 600 .env.local
npm run dev
Open localhost:3000. Keep .env.local private: it is excluded from Git and deployment uploads. Never prefix the key with NEXT_PUBLIC_; it belongs on the server and is not sent to the browser.
npm run typecheck
npm test
npm run build
npm start
The complete input-to-result flow has been exercised with real Gemma calls and live NWS data, including a fresh-clone production run. The verification record documents evidence and limits.
For a running server, npm run smoke:live exercises extraction → forecast → schedule → bilingual briefing and checks that tampered results are rejected. It makes real model calls. Set the existing HEATSHIFT_BASE_URL environment variable to target another deployment.
npm run evaluate:live runs five synthetic work orders three times each against the real model using .env.local. For a focused case, use npm run evaluate:live -- --case=EVAL-02. The runner spaces requests and stops on provider quota errors. Both commands write evidence under the ignored work/ directory.
Kruu is a Next.js Node application. On Vercel, link the intended project, select the Next.js framework and configure GEMINI_API_KEY as an encrypted server environment variable. .vercelignore excludes local credentials. Use the public production URL for the demo; preview deployments can remain protected.
All changes to main, including administrator changes, go through a pull request. CI must pass type checks, focused tests and a production build before merge. Keep branch protection enabled.
Hack for Humanity Miami · Track 01 — Four Forces, One City · Fire
Kruu serves Miami outdoor landscaping crews and their supervisors, with a resident heat guide extending access to local forecast information and cooling-site listings. The submitted entry includes Track 01 — Four Forces, One City, Best Open-Source AI Project and Best Use of Gemma 4. Submission was confirmed on October 10, 2026 at 3:24 PM Eastern. Judging and eligibility remain subject to the official challenge guide.
| Contributor | Contribution |
|---|---|
| Aladdin | Engineering and Integration |
| Faye | Creative Direction, Branding and Demo Story |
| Nelson | Community Impact and Worker Needs |
| Nina | Evaluation and Product Insight |
The submitted team is Aladdin, Faye and Nelson. Nina is credited for her evaluation contribution. Her illustration in the end card is a placeholder avatar.
Build plan · Worker scenario · Evaluation plan · Demo runbook
Original project code and documentation are available under the MIT License. Gemma 4, dependencies and source data retain their respective licenses and attribution requirements. The public source and setup instructions let you inspect and run Kruu with your own API key.
Calendar export follows RFC 5545 for UTC timestamps, text escaping and UTF-8 line folding. Google Calendar uses its event-template URL; no calendar account access or automatic event creation is required.