Selection basis: GitHub star surges over the past ~24 hours. The primary metric is recent growth (findarepo’s measured 7-day and daily deltas, July 13 snapshot); cumulative stars are secondary. Large “rank-holding” repos whose growth has stalled are excluded.
MadsLorentzen/ai-job-search — +15,000 ⭐/week (≈ 22,000 ⭐ total)
github.com/MadsLorentzen/ai-job-search
A local-first AI job-application framework built on Claude Code: fork the repo, fill in your profile, and Claude evaluates job postings, tailors your CV, writes cover letters, and preps you for interviews — all on your own machine. Its defining idea is a drafter–reviewer split: one agent drafts while a second Claude agent researches the target company and critiques the draft, catching missed keywords, weak framing, and generic phrasing before the drafter revises. It never fabricates skills or experience. The project cooled off in early July (daily growth fell from +3,400 to +622) and then re-exploded this week to +15k stars/week, taking findarepo’s No. 1 “top mover” slot — propelled by the real-world story of a geophysicist who built it to run his own job hunt after being laid off.
- Drafter–reviewer separation for higher-quality, self-critiqued output
- ATS keyword-coverage check by extracting the rendered PDF’s text layer
- Iterates LaTeX so the CV is exactly 2 pages and the cover letter exactly 1 page
- Strict no-fabrication rule for skills and experience
Practical use: a good fit for recruiting teams and career coaches who want to validate and demo bulk posting screening and ATS-optimized resume generation locally.
Tags: #JobSearchAutomation #ClaudeCodeWorkflow #ATSOptimization #CVTailoring #DrafterReviewer
JustVugg/colibri — +2,000 ⭐/day (≈ 7,900 ⭐ total)
A pure-C, zero-dependency inference engine that runs GLM-5.2 — a 744B-parameter MoE model — on a consumer PC with just ~25 GB of RAM. It treats VRAM, RAM, and disk as a single managed memory hierarchy: only the dense parameters (~9.9 GB at int4) stay resident in RAM, while the 21,504 routed experts (~370 GB total) are streamed from disk on demand via a per-layer LRU cache. Measured performance runs from 0.05–0.1 tok/s cold on a dev box (WSL2, 25 GB RAM, ~1 GB/s NVMe) to about 1 tok/s on an Apple M5 Max with 128 GB RAM and a fast SSD — so the appeal is less raw speed than the fact that a frontier-scale open-weight model runs at all on commodity hardware. That, paired with pent-up demand for local execution of huge open-weight models, is driving roughly +2,000 stars a day for this brand-new project.
- MLA attention with a 57× compressed KV-cache
- Native MTP speculative decoding (2.2–2.8 tokens/forward at int8)
- AVX2 integer dot-product kernels and async expert read-ahead
- A learning cache that auto-pins hot experts, plus an offline FP8→int4 converter
Practical use: useful for teams that want to experiment with frontier-scale open-weight models on-prem or at the edge without a GPU (proof-of-concept work).
Tags: #MoEInference #ExpertStreaming #LocalLLM #INT4Quantization #SpeculativeDecoding
iOfficeAI/OfficeCLI — +7,300 ⭐/week (≈ 16,000 ⭐ total) · 2nd day running
github.com/iOfficeAI/OfficeCLI
An open-source Office suite purpose-built for AI agents to read, edit, and automate Word, Excel, and PowerPoint files. It ships as a single binary that needs no Office installation, and on install it auto-injects an “officecli” skill into every coding agent it detects — Claude Code, Cursor, Windsurf, GitHub Copilot and more — handing document control to your agent in one line of code. Because agents can manipulate tables, formulas, and formatting accurately without parsing XML, building document-automation pipelines becomes far simpler. After +6.4k stars/week yesterday, it accelerated to +7.3k/week today, so it clears the “still surging” bar and appears for a second consecutive day.
- Built-in engine renders .docx / .xlsx / .pptx to HTML or PNG with high fidelity
- Path-based addressing: every element has a stable path, so agents navigate without knowing XML namespaces
- Built-in formula and pivot engine that auto-evaluates 350+ Excel functions on write
Practical use: well suited to back-office automation that generates, edits, and validates high-volume Office documents — contracts, financial reports, proposals — with agents.
Tags: #DocumentAutomation #DocxXlsxPptx #SpreadsheetEngine #AgentSkills #SingleBinary
Notes
Excluded today for decelerating or plateauing growth: DeusData/codebase-memory-mcp (+3.8k → +3.6k/week), Panniantong/Agent-Reach (+4.1k/week, flat, large at 55k), calesthio/OpenMontage (+3.9k/week, flat), stablyai/orca and diegosouzapw/OmniRoute (still under the prior plateau call), and esengine/DeepSeek-Reasonix (continued surge unconfirmed). Perennially excluded rank-holders include mattpocock/skills (168k), obra/superpowers (253k), NousResearch/hermes-agent (214k), firecrawl (150k), and JuliusBrussee/caveman (89k).
Daily AI Brief 
Leave a comment