Kaveen Himash

Always trying to use technology for the betterment of the people.

[Kaveen Himash @kaveenhimash]

@mmunozpz @invokebio Thanks for the tip, María! I just followed @invokebio. Looking forward to the updates!

[Kaveen Himash @kaveenhimash]

@TXN864 @Dahacker70 It’s a good option if someone wants to advertise a product but doesn’t have an established social media account to do so.

[Kaveen Himash @kaveenhimash]

@12matve You should definitely use it to push a mission to Jupiter next.

[Kaveen Himash @kaveenhimash]

@VaibhavLab What do you usually listen to when doing work?

[Kaveen Himash @kaveenhimash]

@kapilansh_twt In my view, I think AI is replacing the people who don’t have the capital to invest in AI.

[Kaveen Himash @kaveenhimash]

@12matve This looks incredible! I'm not very familiar with the Salyut space station, satellites as a whole, or SpaceFlightSimulator, but I can tell a lot of work went into this. Great job recreating it!

[Kaveen Himash @kaveenhimash]

@IsaiahDoesBiz Yes Claude is the corrector

[Kaveen Himash @kaveenhimash]

Agentic coding burns tokens on work that doesn’t need a frontier model: reading big files, boilerplate, mechanical edits.

So I built Puppet Master, a research concept for cutting that cost. It’s an MCP server where Claude plans and reviews, and a cheaper worker model does the grunt work. The worker can be local (llama.cpp, Ollama, LM Studio, vLLM) or hosted (OpenRouter).

Repo: Repo link in the replies

The guardrails are enforced in code:
• Reads only return claims whose quoted lines were found in the file
• Writes are staged as a patch and run through your tests in a throwaway worktree
• Nothing reaches the repo unless the check passed

Results on a 27-task benchmark, on an M3 Pro with 18 GB running a 2-bit 40B MoE model:
• about 74% fewer Claude tokens (estimated from character counts, not billed tokens)
• 85-90% of reads correct, 86% of writes passing their checks

Accuracy rises with a better worker model. I only tested a heavily quantized one, so treat these numbers as a floor.

Early, MIT licensed, and only tested end-to-end on llama.cpp on macOS. Feedback welcome.

[Kaveen Himash @kaveenhimash]

@b1rkh0ff Let’s gooo

[Kaveen Himash @kaveenhimash]

@DataAnalyt2539 That’s a really good point. I’ll have to try that out.

[Kaveen Himash @kaveenhimash]

@rahuljanagouda That's a great point. I'll look into adding this to the next run.

[Kaveen Himash @kaveenhimash]

Honestly, no. Quote validity is checked in code on every read. Completeness is only checked in my benchmark, by keyword match, so a real-but-partial answer would get through in normal use. I want to add a separate completeness check next, something like requiring each part of the question to be backed by a cited line. Any ideas on how you’d do it?

[Kaveen Himash @kaveenhimash]

@harjjotsinghh Thanks, that’s a really useful data point. It matches what I’d hope for. If the tests are the gate, a smaller model’s mistakes get caught.

[Kaveen Himash @kaveenhimash]

Good question. None of the wrong reads were paraphrased quotes, since those get dropped before Claude sees them. Some were my grader being picky (“rounding up” instead of “ceil”). One was a real miss: real quotes, but half the question unanswered. So yes, different fixes, and checking that a quote answers the question is what I’m working on next. Only 20 questions, so take it loosely.

[Kaveen Himash @kaveenhimash]

Update: I re-checked my read numbers. After fixing a picky grader, reads come out at roughly 95-100% (20 questions, keyword-graded), not 85-90%. Also, tiny jobs aren’t worth delegating. The savings come from big files. Better worker models should help, but I’ve only tested one heavily quantized one.

[Kaveen Himash @kaveenhimash]

Repo link: https://t.co/dpHmu2oaHQ

[Kaveen Himash @kaveenhimash]

@chasmmmmmmmmmmm If this is an issue, then there must have been a time when using a compiled language or assembly language instead of machine code was discouraged.

[Kaveen Himash @kaveenhimash]

@heart_ Is it possible to make the app fit the screen at the OS level?

[Kaveen Himash @kaveenhimash]

@_zer0day__ Great job 👏