adeel akhani

i built a free way to index and search any docs locally (cli or mcp) with a memory layer runs entirely on your machine via @ollama

3D reconstruction stands as one of the most prevalent problems in the CV space. Made a NeRF from scratch to learn more. Optimized across 2× NVIDIA T4 GPUs with F16 mixed-precision with @runpod. Peeked @ PSNR of 22, video pinned below!

@tanmaymshah @runpod incredible work man

@leonardomontesq 💪💪

him

@ollama note: benchmarks sent the full page tree every time. scoping your index to a specific section url shrinks that tree a lot and helps. there's a how_to_use mcp tool that tells agents exactly how.

@ollama wanted to know which local model handles tree navigation best

i built a free way to index and search any docs locally (cli or mcp) with a memory layer runs entirely on your machine via @ollama

@ollama phi4-mini default, best speed/accuracy tradeoff. gemma4:26b noticeably better if you have the ram.

@ollama the root issue is the task combines 3 things at once: scan 400 semi-repetitive lines, semantically match a query to a page name, and return constrained json. small models can do each individually, they just fall apart when you need all 3 together.

@ollama all 7-14b models scored the same. the task: read 400 node paths, pick the right ones, return valid hex ids as json. most would ignore the json instruction and just answer the question from training data. some hallucinated ids and tuning temperature, num_predict didn't help.

@ollama indexing builds a section tree from the url structure and headings. on search a local llm reads that tree and picks the relevant sections. hybrid vector + bm25 runs inside just those sections. another llm pass reranks and filters. tree reasoning also gets semantically cached.

@ollama 6 cli/mcp abilities: index any doc site. search your docs in natural language. inspect the section tree localdocs built. manage multiple indexed sources. swap the reasoning model. control the memory layer. agents also get a how_to_use tool to use efficiently.

@ollama free, local, no api keys. cli and mcp. install: npm install -g @adeel712/localdocs github: github.com/adeelakhani/lo… npm package: @adeel712/localdocs" target="_blank" rel="nofollow noopener">npmjs.com/package/@adeel… pull phi4-mini and nomic-embed-text in ollama and you're setup. mcp instructions and various other details are in the readme

Built a PM tool over the weekend as an AI-first product designer. PMs drown in data assembly, not actual thinking. Me and Claude Code built Northstar. Risk flags, insight cards woven into your board before you even ask. PMs aren't paid to prep. They're paid to think ahead.

@leonardomontesq 👀

@adeel_712 bro teasing

why buy a super expensive computer and not make full use of these modern day local llms

@owengretzinger keeps asking me to put out a project updates daily but I'm too lazy to navigate through linear's complex UI So I made my own app where I can push updates , tickets and agents. all powered by fff from @neogoose_btw

@tanmaymshah @sportynest we got motion here🤩

@sportynest @adeel_712

Where is the tech community in Toronto and how do I join it

@owengretzinger love it🙏

chatgpt 5.4 vs gemma 4 😬

We're introducing Cursor Automations to build always-on agents.