ProofPulse
Interactive proof-coverage analysis. A three-valued model classifies each spec element as essential, auxiliary or irrelevant to the proof, which exposes vacuous proofs and unnecessary preconditions. Evaluated on 252 programs.
Research prototypes from my Ph.D., the open-source verification ecosystem I work in, and the hardware and university projects that came before.
Interactive proof-coverage analysis. A three-valued model classifies each spec element as essential, auxiliary or irrelevant to the proof, which exposes vacuous proofs and unnecessary preconditions. Evaluated on 252 programs.
Dafny programs analysed for vacuous proofs and unnecessary preconditions, with unsat-core minimization sharpening the results.
The Dafny Assertion Inference System. When Dafny gives up on a proof, DAISY combines LLM predictions with the verifier's error messages to propose the helper assertions that are missing, including several at once. It comes with an extension of DafnyBench whose programs have had assertions removed. It repairs 63.4% of programs missing one assertion and 31.7% missing several.
Generates tests from the specification alone, without looking at the implementation. It uses DNF equivalence-class partitioning and boundary-value analysis, with Z3 synthesizing inputs and expected outputs. The tests kill 93.9% of mutants.
Ranks the statements most likely to cause a failed verification using counterexamples from several traces. Evaluated with MutDafny mutants of DafnyBench programs.
An open-source tool that automatically fixes problems reported by software verifiers, starting with Dafny. It is a collaboration between INESC TEC, INESC-ID and Carnegie Mellon. DafnyFix, which repairs programs with a single transformation (APR Workshop @ ASE 2026), is part of this work.
A Dafny plugin that asks an LLM for lemmas Dafny cannot find on its own, and for a calculational proof when a lemma does not verify automatically.
A code editor built on command sourcing, with an AI coding agent inside it. Every edit, whether you or the agent made it, is a small undoable command, so a session's history is a tree you can navigate instead of a linear undo stack. It runs as a Tauri + React desktop app, a Ratatui terminal UI, or headless CLI agents for benchmarks.
View on GitHub ↗Forks I use for experiments, benchmarks and tooling. Live activity is on the Software page.
Verified Rust for low-level systems code.
RustVerified Verus solutions to HumanEval problems.
RustMutation testing for Dafny programs.
C#Test generation for Dafny programs.
DafnyEquivalence-score benchmark of over 2,000 verifiable Dafny programs for evaluation without ground truth.
PythonLaurel: LLM-generated Dafny assertions, evaluated on DafnyGym.
JupyterDafnyBench, a benchmark for formal software verification.
DafnyFork of the Dafny compiler and verifier.
C#Boogie, the intermediate verification language Dafny compiles to.
C#Synthesizing Verilog programs, where formal methods meet hardware.
VerilogI led verification for a new generation of CSI-2 camera controllers for automotive use. I built constrained-random UVM environments that model the design under test, wrote temporal assertions derived from the MIPI specifications, and used formal verification (VCFormal) on the parts of the design that needed stronger guarantees. I hold an ISO 26262 certification for automotive functional safety.
A 40-person university project that built a long-range wireless mesh of wind sensors. I led the 10-person team that implemented the physical layer on software-defined radio.
A web app that shows research papers as a graph and clusters them by similarity. I built the clustering that sets the distances between nodes.
For my M.Sc. thesis I used dimensionally aware genetic programming to evolve human-readable dispatching rules for the job-shop scheduling problem.