Rak is a general-purpose programming language built mainly for cybersecurity, OSINT and systems work. It's written in Rust and licensed under Apache-2.0.

• Hex-first. Hex is a first-class type, with a static type checker, generics, traits, pattern matching (including binary byte patterns), macros and closures.
• Two backends. There's a tree-walking interpreter and a bytecode VM that runs about 6x faster.
• Forensic structs. binstruct gives you a decoder and an encoder from one declaration. Decoded values carry evidence<T> provenance tags, so report(...) can produce chain-of-custody-cited findings.
• Security stdlib. It covers DNS/TLS/PCAP parsers, raw socket packet forging, WHOIS, certificate-transparency subdomain enumeration, a YARA-lite scanner, native crypto (AES-GCM, Ed25519, HMAC), memory-mapped files and FFI.
• Self-contained. It can build standalone executables and native GUI windows, and it ships a SQL server and a Rak interpreter, both written in Rak.
• Tooling. It has the oyvey package manager, an LSP, a DAP debugger, a formatter and linter, a REPL, a Tauri + Next.js IDE and a VS Code extension.

Latest: v0.7.2 · Windows & Linux · Repo · Website

Senlight AI is a research project on Artificial Emotional Intelligence (AEI): an LLM that tracks its own emotional state and reasons about moral context.

• Affective attention. A decoder-only transformer whose emotional state (VAD: valence, arousal, dominance, anchored to Plutchik's eight emotions) biases the attention logits directly. Emotion changes token probabilities at the arithmetic level instead of through prompt wording.
• Intent router. A dedicated crisis-detection head sits outside the main softmax. The model mirrors a user's own emotion but not third-party emotion, and not crisis.
• Sycophancy harness. An eval suite measures whether the model caves to false premises or pushes back.
• Two planned models.
  • Senlight Elafry (8B): on-device edge model for real-time affective dialogue.
  • Senlight Varys (70B): heavyweight model for deeper psychological analysis and moral reasoning.
• Implementation. The PyTorch training and eval stack has 305 tests. A Rust (candle) inference engine has cross-language parity checks against PyTorch.

Status: early research. The architecture, training stages and tooling are implemented, but no model has been trained yet, so there is no checkpoint.

Repo · Website