Satyawan Singh
AI/ML engineer who builds custom models, not wrappers.
New to Contra
Satyawan is ready for their next project!
I just launched cargo-vouch, an open-source Rust tool that proves a function can't panic. No tests to write, no annotations to add.
Point it at a .rs file, and it builds a Kani proof harness for every function, then checks every input in bounds. Each function comes back with one of four verdicts:
š“ BUG, with the exact input that triggers the panic
š” UNGUARDED, safe except at type extremes like i32::MAX
š¢ VERIFIED, provably panic-free
ā³ INCONCLUSIVE, when it cannot decide
Why I built it: AI writes Rust that passes cargo test, then panics on an edge input nobody tried. For example, numbers.iter().sum() looks fine until something hands it [i32::MAX, 1].
I ran it on 11 functions GPT-4o wrote and graded. GPT-4o's own grading got 4 of them wrong: it passed three that overflowed and failed one that was fine. cargo-vouch called all four correctly, with the triggering input for each real panic.
The honest limit: it is built for loop-light code. Parsers and loop-heavy functions return INCONCLUSIVE, never a false pass. I would rather say that than oversell it.
If you write Rust, I would like your feedback.
Repo - https://lnkd.in/eHirVFQq
(https://lnkd.in/eHirVFQq)cargo install cargo-vouch
#rustlang (https://www.linkedin.com/search/results/all/?keywords=%23rustlang&origin=HASH_TAG_FROM_FEED) #opensource (https://www.linkedin.com/search/results/all/?keywords=%23opensource&origin=HASH_TAG_FROM_FEED) #formalverification (https://www.linkedin.com/search/results/all/?keywords=%23formalverification&origin=HASH_TAG_FROM_FEED)
0
5
Stigmergic consensus for multi-agent / multi-model systems.Ā A learned, stateful alternative to majority vote and self-consistency: agents deposit weighted signals into a shared pool, a short truth-discovery loop converges on consensus, and cross-runĀ pheromoneĀ memory teaches the swarm which agents to trust. Pure Python, zero dependencies, provider-agnostic.
0
14
AI/ML Engineering Custom Models, Agents & Systems
0
0
PRAXIS is my MSc dissertation project at the University of Leicester: an end-to-end deep-learning system for voice-based biomarkers, designed and built from scratch.
I architected and implemented the full pipeline: a self-supervised speech representation model plus a downstream biomarker system, and contributed multiple novel technical and methodological advances, documented in a 15,000+ word dissertation with full research rigour.
What I did:
Designed a custom deep-learning architecture for speech/audio
(self supervised learning + domain adaptation)
Developed several original contributions in [voice-biomarker modelling/representation learning] (kept high-level to protect the method)
Built a reproducible, rigorously evaluated pipeline end-to-end: data ā model ā results
0
53