Curated Open-Source AI Repositories

“I ran paper-trading research summaries for a Surat client with TradingAgents capped at end-of-day reports. Kronos as a market-native foundation model is the forecasting backbone I would eval next for that desk — language of markets instead of general chat bolted onto prices. I would keep it read-only research, never live execution, same gate as before.”

“For IndiaAI mission labs & Gujarat universities — train small BharatGen-style models on ₹65/hr GPUs without closed APIs.”

“I champion local-first from Junagadh (offline 70B, Pi 5) — FckSignups is the directory that keeps my clients off SaaS signup traps. I pull its in-browser tools into my stack docs when I spec no-signup fallbacks for Gujarat SMEs on tight data.”

“I filter 26% vulnerable skills via SkillSpector — skill-flare-discover adds multi-registry search + security labels before install. Its finder is the front-door I recommend to Gujarat teams before any skill hits Junagadh ledger.”

“I catalog awesome-claude-skills with 72k stars — agent-skills-cli scales that to 50k installable skills in one command, solving the discovery bottleneck I hit monthly. Its one-command install is the UX I want for Junagadh SME onboarding.”

“I curate 103 repos for deepakbagada.in/repos — Awesome-AI-resources is the upstream list I cross-check for blind spots before Sep 8 publish. Its Java-heavy curation balances my Python/TS bias for Gujarat client stacks.”

“I maintain 41 skills where docs rot kills agent accuracy — openwiki auto-generates AI-consumable docs so my MCP tools stay discoverable. Its agent-first docs format is what I enforce before any skill enters my Junagadh catalog.”

“I benchmark Phi-4 Mini 3.8B at 300 tok/s on 3GB where every megabyte decides deployability — needle compresses a foundation model to 14MB for wearables and phones, an order beyond my Pi 5 edge work. Its tiny-device target is the endgame of my on-device thesis for 4G-drop Junagadh field agents.”

“I point Junagadh interns here before touching my client stack. Rebuilding tokenizers and retrievers by hand removes magic. Two interns shipped MCP tools within a month after this.”

“I query this radar from my lab before adding anything to this library. Velocity plus the MCP-skill taxonomy beats raw trending for agent work. I wired its MCP server into Cursor for morning scans.”

“I feed Gujarat SME quotations, invoices, and spec sheets into pgvector from Junagadh, and MarkItDown sits at the front of that pipe. A Rajkot RFQ inbox with mixed PDF and XLS attachments went from manual copy-paste to indexed with 42ms P95 HNSW search. I still hand-check scanned Gujarati PDFs — OCR misses cost me one wrong line item in April.”

“I point Junagadh interns here before they touch my client stack — rebuilding one single-file RAG starter removes magic faster than any lecture. Two interns shipped MCP tools within a month after working through the retrieval examples. For client work I vendor exactly one pattern per build and pin model versions.”

“I run a 90-pattern blocklist as an automated audit gate before anything goes live from Junagadh, so a standalone slop remover slots into my pipeline as first pass. It flagged three throat-clearers in one founder-story draft last week that my own pass missed. Final sign-off stays human — no filter can verify a P95 number or a rupee price.”

“I point Junagadh interns at ai-engineering-from-scratch before touching client work, and this is the missing second textbook — design principles plus engineering practice with code per chapter. Rebuilding planners and handoffs by hand removes the magic that causes production surprises. Two of my interns would read this before their first MCP tool.”

“I run AEO plus SEO for Gujarat SMEs where Rs 8k monthly SaaS seats eat the retainer. A self-hosted Semrush-shaped alternative I can point at deepakbagada.in citations plus AI Overview tracking is the tool my 90-day ledger wants. I would trial it against my PAA-mapping workflow for one client quarter before moving spend.”

“I route bulk batches to Sol plus GLM Flash plus Kimi on rupee math from Junagadh, and free-tier aggregation is the missing bottom rung for eval replays. Thirty-four providers behind one gateway means nightly promptfoo runs stop touching paid quotas. I would sandbox it for eval traffic only — never client money paths — with per-provider caps logged to the ledger.”