I build AI tools. Then I prove them.
B.Tech CSE at SRM University AP, class of 2029. Founder of Model D and Face Law Labs. Contributor to Graphify's core and claude-mem.
By one estimate, models will train on datasets as large as all public human text sometime between 2026 and 2032. After that, more of their training data will be written by other models, and someone has to check it. I build small, exact systems for that job. They run at no token cost.
The examples don't pin one program down.
The clip follows no law it can stand behind.
The text doesn't settle it either way.
Below a proven light limit, it never guesses.
PLUMB
A model writes a few examples. PLUMB compiles them into a program with no model inside, and the program labels as many new inputs as you want for free.
"the quick brown fox" → "TQBF" "hello big world" → "HBW" "plain words here" → "PWH"
def f(s):
return ''.join(([w[0] for w in ((s.upper()).split()) if w]))[3, 1, 2] → 3
The length, the largest number and the first number all fit. One example can't tell them apart.
Real output from PLUMB 0.2.0.
new labels from 95 programs, with 0 errors against their own programs. All 95 reproduced the examples Claude Opus 5.5 wrote for them.
against Opus 5.5 on 100 fresh machine-made tasks, right minus wrong.
per million labels at 1,000 labels per task, against Opus writing every label.
What it can't do yet. On rules written by people, Opus got all 50 right and PLUMB answered half.
AM Math 3D
AM, a core reasoner, turns a video into exact 3D numbers and locks the shot you ask for before anything is generated. Every painted frame is then measured against the lock.
median gap from the locked plan, worst 3.4 px over 73 frames. The whole shot cost about $1.05. AM's own part took 166 s on a laptop and 0 tokens.
What it can't do yet. A 15-second ad failed: the model returned 4 seconds and ran about twice as fast. On the mouse, AM's core answered CANNOT on the clip's motion, so the jump's math was written by hand from its measurements.
Detect Law
It reads two small open models and decides with integer arithmetic, offline on a laptop. RAID's own bot scored it on 672,000 hidden texts.
of AI text caught at 5% false positives, on the largest AI-text detection benchmark.
What it can't do yet. As far as I can find, this is the highest zero-shot score on the board, but detectors trained on AI text still rank above it.
Model D
It keeps what a coding agent learns about a codebase on disk and tells it where to look in every new session.
cheaper over Django's 288 fixes from 2022, at the same accuracy: $12.32 against $25.72, with 252 vs 253 right.
What it can't do yet. An outside agent's tests on unseen codebases ranked it first but measured a smaller gain, 1.58× fewer tool calls.
Face Law
Counted integer tables replace a trained network. No GPU, and no training beyond counting.
of owners accepted at 1 stranger in 1,000, in a NIST-style test on Extended Yale B. The best classic method reached 84.0%.
The whole model.
Fastest comparison, up to 19 µs.
One CPU core, integer look-ups only.
Comparisons passed NIST's own test driver.
What it can't do yet. So far it is tested on lab faces only, the NIST entry is not submitted yet, and a photo of the owner can still open the phone unlock.
of 176 bugs fixed by a small model, against 30.7% alone. A model-free engine that sits beside a language model, mutates its code token by token and lets the tests judge every variant. It did nothing for problems a strong model already solves.
contest problems on LiveCodeBench v6, none lost, at about $0.20 per extra solve. It makes a model check and fix its own answers before they ship. On rule-finding questions it was 53 to 122× cheaper than Gemini alone, with more wrong answers, 8 vs 1 to 5.
Open source
Four commits in its core query code, credited in the release notes. Strongest matches now print first, which put the fix file in the top 5 on 63 of 114 real Django bugs, up from 53. --seed lets an outside ranker choose where a search starts. Three more changes are in review.
My fix to its search service made plain-English questions work. On LOCOMO's 1,536 questions, recall@10 went from 0.001 to 0.534. A sync speed-up, 131 ms to 2.6 ms on a 3.9 MB page, is in review.
Two Detect Law versions are on the leaderboard of the ACL 2024 benchmark. I also proposed tags that disclose whether a detector trained on RAID's data, now in review.
10.5281/zenodo.23937716
10.5281/zenodo.23937763
With Folarera Ayobami Kassim. 10.5281/zenodo.23861718
Software record. 10.5281/zenodo.23289068