Solo, self-taught software engineer building AI systems from first principles - from OS kernels and GPU runtimes up through language models trained from a blank directory.

Currently building: aide-sovereign-workbench - AIDE, an offline-first AI development environment: local model runtime, agentic coding loop, an evidence-based verification harness, bring-your-own-key provider routing, and hybrid retrieval. Every subsystem ships behind typed contracts, tests-first coverage, and CI gates, and the whole environment is designed to run fully offline on consumer hardware.

• Small language models from absolute zero. Tokenizer training, custom architectures, pretraining pipelines, post-training (supervised fine-tuning, distillation, preference optimization), evaluation gates, quantization to GGUF, and local serving on consumer GPUs (6 GB VRAM class). Multiple model families designed and built from scratch - the training pipeline itself is treated as a first-class product with stage gates and adversarial evaluation.
• Systems engineering at every layer. Process and memory models, driver and runtime constraints, service lifecycle on Windows and Linux, deterministic builds, port and process management. Comfortable working below the abstraction layer most tools pretend does not exist.
• Security-first architecture. Threat modeling, least privilege, fail-closed defaults, OS-level secret handling, and manipulation-resistant agent design informed by social-engineering research.
• Verification as culture. Evidence or it did not happen: real test batteries over smoke checks, honest limitations documented per release, adversarial suites, reproducible evidence artifacts.

Solo and long-horizon. Every capability ships through the same loop: research from primary sources, design doc, implementation, full verification battery, public evidence notes. Sponsors are not funding ideas - they are funding a demonstrated cadence, sustained week over week across two active fronts: an offline IDE and from-scratch models.

• Dedicated compute for larger training runs - the current work proves the entire pipeline on a GTX 1060; scale is the remaining variable
• Hardware for a broader device test matrix
• More research-to-release cycles, faster

If you believe in independent engineering that proves everything it ships:

github.com/sponsors/AnonymousNomad

Contact: open a discussion or issue on any active repository.