Satyawan Singh
I build systems that prove things: verification for AI output, formal checks for code, receipts for machine decisions.
Currently Velkron, the verification layer for AI in regulated industries. Before and around that: a formally tested blockchain, a clinical speech model trained from scratch, and a set of open-source libraries that all circle one conviction: machine output should be provable, not trusted.
Everything below is real, dated and checkable.
ADVISORY
full details →Independent review of where a model's output is trusted without being checked, and what a real verification layer would look like.
Rust and z3-based proofs: panic-freedom, safety envelopes, and other properties that should be proven, not tested for.
EU AI Act and FCA-adjacent work: provenance-gated systems, refusal-first design, and audit-ready documentation.
For teams building AI in healthcare, finance, or other regulated industries. Building Velkron full-time, so this is a small number of engagements alongside that, not a full-time practice.
See a live safety proof ↓SIGNED CLAIMS
click verify on any cardThe verification layer for AI in regulated industries. Deterministic compliance checks, tamper-evident receipts, proof of inference.
Rust tool that proves a function can't panic, built on Kani. Four honest verdicts; INCONCLUSIVE is never a false pass.
Four zero-dependency Python libraries, one idea: check the output, don't trust it. Conformal abstention, grounding, verifier-gated search, agent consensus.
A formally-verified runtime shield that proves a robot control policy stays inside a safety envelope. Every action is checked by the z3 solver, demonstrated on a MuJoCo arm.
Trust-scoring for clinical AI: when should a medical model's output be believed, and when should it abstain?
NHS GP scribe with a 119M-parameter Conformer ASR model trained from scratch for UK clinical audio.
RegTech retrieval for FCA-regulated firms, with provenance-gated citations and a refusal-first answer gate.
A 140-crate Rust L1 blockchain with 12,950 passing tests and novel consensus-verification primitives. Build complete; research continues.
RUNNING DEMOS
Fig. 1 — Each leaf is a claim; the root changes if any one of them does.
Fig. 2 — Drag the red zone onto the path, then prove the shield keeps the arm clear.
WRITING
view all →- Why AI needs receipts↗ velkron.xyz
In a world where models produce confident-sounding output, the only real defense is proof, not trust. On what a tamper-evident receipt for machine output actually requires.
- The most important sentence in the EU AI Act↗ velkron.xyz
One requirement buried in the regulation will reshape how compliant AI systems get built. Most teams haven't read it yet.
- Refusal is a feature↗ velkron.xyz
When a system says it doesn't know, that's not a failure state. It's one of the few honest safety mechanisms available, and most designs actively train it out.
SHIP LOG
- 2026-07-23Launched velkron.xyz and console.velkron.xyz. Applied to Hub71. Shipped verified-safety-shield.
- 2026-07-21Launched cargo-vouch on crates.io.
- 2026-07-13Published the open-source verification quartet.
CONTACT
Building Velkron full-time. I advise a small number of teams on AI verification, LLM systems and applied ML.
intent
from
message
manifest
- timestamp
- —
- intent
- —
- payload_length
- 0
- payload_hash
- —
[UNSAT — incomplete]
payload_hash is a real SHA-256 of what you type, computed in your browser via crypto.subtle — nothing is sent until you commit, which opens a pre-filled email to satyawansinghinuk@gmail.com.
CONSOLE
A command-driven way to browse the same real data above. Type help to see what's available.