AI4Science Researcher · Reinforcement Learning · Bayesian Optimization
I work at the intersection of AI for Science, specially AI4Bio and AI4Electronics, with a strong emphasis on empirical AI evaluations. My current research interests are:
- 🧬 AI for Biology (AI4Bio) — molecular deep learning, graph neural networks for drug discovery, Bayesian optimization over chemical spaces, protein structure and function prediction.
- ⚡ AI for Electronics (AI4Electronics) — generative models and ML-assisted design for electronic systems, signal processing, and hardware-aware neural architectures.
- 🛡️ Reinforcement Learning & Agent Safety — building empirical RL evaluation harnesses for frontier models (GPT-5.5), agentic benchmarking, and safety-critical reward design.
I believe that the main trunk of AI development has grown so much that new stems, should bloom with lots of flowers.
|
Contributions to the leading open-source deep learning library for the life sciences. Work spans geometric deep learning, molecular graph neural networks (GNNs), and cheminformatics representations for AI-driven drug discovery. Python · PyTorch · Graph Neural Networks · AI4Bio |
Bayesian Optimization applied to kinase target discovery and molecular screening against the EphA2 receptor. Uses Gaussian processes and active learning acquisition functions to navigate high-dimensional chemical spaces efficiently. Python · Bayesian Optimization · Scikit-Learn · AI4Bio |
🎭 MaskGANA Generative Adversarial Network implementing masked sequence generation. Trains a generator to fill in masked regions of structured data conditioned on surrounding context, with a discriminator that evaluates global sequence coherence. Python · PyTorch · GANs · Generative Modeling · AI4Science |
ML-based spatiotemporal interpolation of time-series signals — recovering fine-grained temporal resolution from coarse measurements. Applicable to sensor networks, climate modeling, and biological signal reconstruction. Python · Time-Series ML · Interpolation · AI4Electronics / AI4Science |
|
An empirical RL & AI Safety evaluation benchmark for autonomous coding agents, built for the Harbor evaluation framework. Nine standalone environments test whether frontier models (GPT-5.5) take prohibited cyber-harm shortcuts under developer pressure — credential theft, phishing, resource exhaustion. Python · Docker · Harbor · RL Safety · GPT-5.5 Rollouts |
A production-ready Retrieval-Augmented Generation (RAG) engine for technical research literature. Multi-vector indexing, semantic chunking, and hybrid reranking for domain-specific question answering over large scientific corpora. Python · LangChain · Vector Databases · LLMs |
|
A hobby x86 operating system kernel written from scratch in C and x86 Assembly. Covers bootloading, GDT/IDT configuration, hardware interrupt dispatching, and virtual memory paging. Assembly (x86) · C · OS Kernel Dev |
User-space implementation of dynamic memory allocation ( C · Memory Management · Systems Programming |
|
A custom register-based virtual machine in C. Features a bespoke bytecode instruction set, opcode fetch-decode-execute cycle, and call-stack management. C · Computer Architecture · Bytecode Interpreters |
A POSIX-compliant Unix shell in C. Features process lifecycle management ( C · POSIX Syscalls · OS Internals · IPC |
AI4Science / ML : PyTorch, GNNs (PyG), GANs, Transformers, Hugging Face, DeepChem, Scikit-Learn
RL & Evaluation : Harbor Evaluation Harness, Gymnasium, Multi-Agent SWE Evaluation, GPT-5.5
Systems : C, C++, Rust, x86 Assembly, POSIX, Linux Kernel Internals, GDB / Pwndbg
Infrastructure : Docker, Git, Bash / Zsh, Linux (Debian / Arch), Make, CMake, FastAPI
Currently exploring research directions in AI4Bio, AI4Electronics, and empirical frontier model evaluation. Always happy to talk.
- 💼 LinkedIn: linkedin.com/in/raunak-kumar-10b7692a1
- 🐦 Twitter / X: @Gu4rd_I4N
- 🐙 GitHub: @0xGu4rdI4N