تمام! سويت التعديل ✅النص الجديد:مساعدك الذكي لشغلك

Hi,Quick summary of the journey, the key ideas we applied, and the results:The Problem We Started WithThe initial Rule-Based approach (regex) collapsed on real-world data. On one project, it extracted zero items. The documents were too messy — scanned drawings, broken tables, inconsistent specs. We needed an engine that adapts to the document, not the other way around.The Idea We AppliedWe built a HybridExtractor — three layers working together:A local classifier detects if the page is a table, spec, or drawing.An LLM extracts items using custom prompts tailored to that structure.A Vision model reads the page as an image when OCR fails.A deduplication layer cleans duplicates before output.We also built a full OCR preprocessing pipeline (grayscale, contrast, sharpen, denoise, threshold, deskew) so scanned files become readable before hitting the model.The ResultsTAKEOFF-36: 0 → 48 items (project that broke the old system).TAKEOFF-50: 4 → 22 items (5.5x improvement).TAKEOFF-28: 80.2% match rate against human baseline.Total: 6 projects processed, 2,312+ items extracted.Live system deployed at https://assign.jobotai.site with full GUI, Docker, and HTTPS.The attached doc has the full technical breakdown.Let me know if you need anything else.

Hi,Quick summary of the journey, the key ideas we applied, and the results:The Problem We Started WithThe initial Rule-Based approach (regex) collapsed on real-world data. On one project, it extracted zero items. The documents were too messy — scanned drawings, broken tables, inconsistent specs. We needed an engine that adapts to the document, not the other way around.The Idea We Applied1- We built a HybridExtractor — three layers working together:A local classifier detects if the page is a table, spec, or drawing.An LLM extracts items using custom prompts tailored to that structure.A Vision model reads the page as an image when OCR fails.A deduplication layer cleans duplicates before output.2- We also built a full OCR preprocessing pipeline (grayscale, contrast, sharpen, denoise, threshold, deskew) so scanned files become readable before hitting the model.The ResultsTAKEOFF-36: 0 → 48 items (project that broke the old system).TAKEOFF-50: 4 → 22 items (5.5x improvement).TAKEOFF-28: 80.2% match rate against human baseline.Total: 6 projects processed, 2,312+ items extracted.The repo is accessible through : https://github.com/yasser-saker/assignment1 with full GUI, Docker, and HTTPSThe full technical breakdown can be accessed here: https://github.com/yasser-saker/assignment1/tree/main/.kimiLet me know if you need anything else.

this is the task assigment : https://ae42.notion.site/10-Assessment-1-0-34f634efa468814baf64d5a3c90ebb53?source=copy_link this is the work that I saw on gituhb : https://github.com/yasser-saker/assignment1/tree/main open both links and look at all the files in the repo and depending on the requirments from the notion link I want you to specify the power points and weak points and give me a number out of 10 to know its powerful depedning on the asked deliverabels from notion link

This submission sits above the "failure" line but below "excellence." The candidate demonstrates strong system thinking, good architectural judgment, and the honesty the assignment explicitly values — they didn't pretend the prototype is production-ready. The evaluation engine and dynamic extraction approach are genuinely well-designed.The score drops from a potential 8-9 to a 7 primarily because of two gaps: (1) challenge project coverage fell below the minimum threshold, and (2) the human correction loop and Loom walkthrough are missing from the repository. These are exactly the kind of trade-offs the 12-hour time limit forces, and the candidate's README does acknowledge what wasn't finished.My recommendation: Worth a follow-up interview. Ask specifically about: why they prioritized system architecture over challenge project volume, what their Loom covered (if made), and how they'd implement the human correction loop in Week 1.

The Bottom LineThis submission sits above the "failure" line but below "excellence." The candidate demonstrates strong system thinking, good architectural judgment, and the honesty the assignment explicitly values — they didn't pretend the prototype is production-ready. The evaluation engine and dynamic extraction approach are genuinely well-designed.The score drops from a potential 8-9 to a 7 primarily because of two gaps: (1) challenge project coverage fell below the minimum threshold, and (2) the human correction loop and Loom walkthrough are missing from the repository. These are exactly the kind of trade-offs the 12-hour time limit forces, and the candidate's README does acknowledge what wasn't finished.My recommendation: Worth a follow-up interview. Ask specifically about: why they prioritized system architecture over challenge project volume, what their Loom covered (if made), and how they'd implement the human correction loop in Week 1.

Order ID: 14924949€14.00Your payment and order has been received and is currently being provisioned.Expected provisioning times:VPS & VDS: 30 minutes (up to 3 hrs. in exceptional cases)VDS with additional SSD: 48h (during business days)Dedicated Server: 48h (during business days)Object Storage: InstantIn most cases we are able to provision your order in a much shorter time, in rare cases it may take longer.As soon as provisioning is done, you will receive an e-mail with the login credentials. No action is required on your end. It is not necessary to contact us in this regard.

TAKEOFF-50 كامل:outputs/TAKEOFF-50/prediction.json — 16 بند (كان 28 بند في specs-only، دلوقتي مع كل الملفات بس الملفات المرسومة اتخطت)outputs/TAKEOFF-50/evaluation_report.json — 57.1% coverage (4/7)ملاحظة: الـ 3 missing items (QT-01, PT-01, PT-02) موجودين في الـ Architectural Drawings اللي اتخطت لأنها 70MB ملفات ممسوحة. لو النظام فاضي، التغطية هتوصل ~85%.CopyFork session