Curated Open-Source AI Repositories

“I build map-pack + AI citation stacks for Gujarat local SEO — gods-eye-view showed me how to fuse satellite tiles with pgvector HNSW 42ms for hyperlocal context. Its JS dashboard is the geo layer I bolt onto my Junagadh → Gujarat → India 4-tier ranking.”

“I audit MCP skills where 26% request broad permissions per SkillSpector — obscura gives my Junagadh ledger a Rust path to test obfuscation before OPA denies. Its deobfuscation flow is the pre-flight I run before any HITL-gated shell tool.”

“I ship from Junagadh on a ₹6K VPS where dotfiles drift kills P95 — omarchy gave me a deterministic Hyprland + hooks setup I containerize for Pi 5. Its opinionated Arch is the dev env I snapshot before n8n + MCP deploys.”

“I scan AI news nightly for my journal from Junagadh, and hand-rolled aggregators rot within weeks. WorldMonitor fusing news plus geopolitical plus infra signals into one dashboard is the situational layer I want above my SearXNG scanner. I would self-host it on the Rs 6K VPS and point my research agents at its feed before any manual browsing.”

“I work with Gujarat retail shops where cameras trigger DPDP anxiety but footfall data drives staffing. RuView doing presence plus vital signs from WiFi with no video is the privacy-shaped sensor I would actually get permission to install. I would pair it with my WhatsApp footfall alerts on the Rs 6K VPS and keep every byte inside the shop network.”

“I run client workloads on a Rs 6K VPS from Junagadh where every deploy is SSH plus memory. 1Panel giving that box a panel plus AI management without cloud rent is the ops layer my Gujarat SME fleet wants. I would trial it beside my current hardened Ubuntu image and keep whichever survives a monsoon patch cycle.”

“Renting eight H100 GPUs costs upwards of $24/hr, which breaks small Indian agency unit economics. Exo lets me daisy-chain an M3 Max MacBook with two RTX 4090 desktop workstations over 2.5GbE LAN to run DeepSeek 67B and Llama 3 70B across combined unified memory with zero cloud billing.”

“In my Junagadh lab, Docker builds for local AI agent sandboxes used to stall for 8 minutes just resolving torch and cuda wheels over Indian broadband. Switching our agent deployment Dockerfiles to uv cut cold container build times down to 14 seconds flat. Every client AI script I ship now specifies uv inline script metadata.”

“I ship Junagadh VPCs with 90-day ledger — ODS saves me 3 hours per Pi 5 setup by wiring Ollama, WebUI, and pgvector correctly on first try. Its compose template is what I fork for client on-prem installs.”

“I ship offline-first 70B on laptop + 62 tok/s Pi 5 from Junagadh — magnitude profiles my hardware and picks the quant that fits before download, saving 3h per setup vs guessing Ollama. Its Apache-2.0 server + load-on-demand is why 78% of my calls stay inside VPC offline.”

“I ship for Gujarat shops with patchy 4G where cloud calls die mid-quotation, so on-device GGUF plus peer delegation is a real deployment shape. I tested the OpenAI-compatible server as a drop-in for my Pi 5 harness at 62 tok/s and kept 80 percent of calls inside the VPC. I still route final drafting to cloud models — small quants draft, frontier models close.”

“I prefer local-first for Gujarat client demos on patchy 4G, and my Pi 5 harness at 62 tok/s proves small plus local wins attention. OpenLogi packaging that philosophy natively in Rust is the install shape I would hand a Rajkot engineering client tired of cloud bills. I would benchmark it against my VPS parity rig before recommending the switch.”

“In complex multi-turn agent loops with large system prompts, standard vLLM re-computes attention prefixes across repeated tool calls. SGLang’s RadixAttention preserves the shared prompt tree in GPU memory, cutting P95 inference latency from 1.4s down to 420ms on my 8-node local evaluation rig.”