Final-semester Computer Science student at Reichman University (graduating February 2027), looking for junior AI/ML and software engineering roles in Israel. I like building systems I can measure: frozen test splits, CI on every change, and an honest list of what didn't work.
Site: yuvgev.com · Email: yuvgev1@gmail.com
- taintwall: a runtime firewall for tool-using LLM agents. A taint-tracking policy blocks side-effecting tool calls (send money, send email) whose arguments came only from untrusted tool output. On 461 held-out AgentDojo attacks against a live model that falls for them, it blocks as many as the ProtectAI detector (55.5% → 13.7%) while keeping 72% of benign tasks working versus 44%. Python, FastAPI, Docker
- Hebrew ToolBench: a Hebrew tool-calling benchmark: 800 items over 52 Israeli-domain tools, scored by a deterministic AST scorer, with 9 models evaluated. The Hebrew calendar is the hardest part for every model, and the "wrong language for a canonical value" failure appears only in the smallest model. Python
- mini-engine: an LLM inference engine written from scratch in PyTorch, with a paged KV cache, continuous batching, prefix caching and n-gram speculative decoding. Its greedy output is token-identical to Hugging Face, and it is benchmarked against vLLM on an L4: speculative decoding gives 4.5–4.7× on copy-heavy work. Python, PyTorch
- iHelp: a reverse help marketplace where verified neighbors compete to help (live). Every permission rule is enforced inside PostgreSQL with row-level security, plus 65 automated tests. Next.js, TypeScript, Supabase
- SubScan: a Hebrew web app that finds forgotten subscriptions in the statements you export from Israeli card issuers and banks, and drafts the cancellation letters. In beta; the code is private. TypeScript
- penguin-forge: a real-time multiplayer world on an HTTP/WebSocket framework I hand-wrote on Node's raw
netmodule, with zero dependencies. JavaScript, Node.js
- Opened ethz-spylab/agentdojo#228, which adds support for current Claude models to AgentDojo (under review).