I am an AI systems engineer and open-source founder in Agra, India, with a CSE (AI/ML) background from VIT. I work across the full stack: Rust kernels and no_std userland, Python agent runtimes, TypeScript interfaces, model evaluation, CI, releases, and product design.

My recurring question is:

How do we give an agent useful autonomy without confusing intelligence with authority?

That is why my projects publish capability boundaries, reproducible evidence, failure history, and limitations alongside the features.

Heliox is an MIT-licensed desktop agent for Windows, macOS, and Linux. It converts natural-language and opt-in multimodal input into validated plans, asks before risky work, and routes actions through permission and verification contracts.

• 157 declared action types backed by 21 executable specialists.
• Text, continuous voice, gesture, gaze-region signals, screen context, and experimental neural input share a priority-controlled interaction path.
• Local and cloud model providers plus text-only planning through officially authenticated Codex and Claude Code CLIs; provider credentials remain with the provider tools.
• Local IDE MCP, secure Air Handoff, calendar/email/allowlisted-SSH integrations, plugin manifests, a repository-governed marketplace, approval gates, rollback planning, and outcome verification.
• Public capability catalog, benchmark artifacts, release feeds, failure notes, and agent-readable documentation.
• The measured guarded CPU-status fast path is 27.229 ms median / 29.476 ms p95 across 100 non-LLM iterations; it is not presented as end-to-end voice, browser, provider, or UI latency.

Current limitation: provider coverage is not the same as independently observed success. Eleven of the 157 actions currently have an independent post-condition verifier; hardware accuracy for microphones, cameras, gaze, gestures, and EEG requires human testing.

FerrumOS studies what changes when an agent acts through an operating-system security boundary rather than only through application-level tools.

• Bootable Rust kernel, ELF loading, real Ring-3 processes, scheduler, filesystem, GUI, and package/runtime services.
• 61-syscall ABI, 41 canonical agent operations, and five permission tiers.
• Every public tool path passes through deterministic policy and a provider-independent predictive screen before execution.
• Public QEMU evidence covers 101/101 focused command paths and 81/81 catalog entries for the recorded audit commit.
• The simulator-backed cyber-physical software tier passes 152/152 deterministic contracts and 32/32 physical-model, robustness, and neural-decoder gates.
• Ring-3 world-model preview measurements span 1.29–1.57 ms run means across H=1…5 with zero measured heap growth in three runs.
• Synthetic neural intent is bounded and fail-closed; physical predictions remain shadow-only and cannot grant actuator authority.

The published fixture reports 81.4% balanced accuracy for rules + JEPA vs 81.2% for rules + a per-action mean baseline. That is deliberately documented as no material JEPA advantage on that fixture—not hidden behind the headline.

FerrumOS is a QEMU/Bochs research OS. The cyber-physical results are software and simulator evidence—not proof of installed third-party simulators or transports, robot execution, broad physical-PC compatibility, hard-real-time behavior, formal safety certification, or validated live EEG control.

Nucle-OS treats molecular storage as a software system with explicit codecs, recovery, indexing, policy, observability, and hardware-provider boundaries.

• Eight-layer engine: synthesis simulation, encoding, ECC, retrieval/indexing, VFS, agent interface, hardware bridge, and block-device abstraction.
• Yin–Yang, ternary, fountain, and direct-packing codecs with biological constraint checks.
• Reed–Solomon error-and-erasure recovery, fountain erasure repair, and partial-order-alignment consensus.
• Durable VFS with versioning, migration, capacity limits, encryption, integrity scanning, tenancy, audit logs, and Prometheus metrics.
• NucleScript compiler, type/effect system, LSP, VS Code extension, package registry, CLI, native playground, and browser WASM playground.
• 729 discovered tests: 728 passed and one ignored in the documented workspace run.

The project simulates and integrates DNA-storage workflows in software; it does not claim that every workflow has been run against a production wet lab.

I am interested in collaborations around agentic operating systems, capability security, assistive multimodal interfaces, world models, compilers, and molecular storage.

• Start with Heliox issues if you want to improve a user-facing agent.
• Start with FerrumOS research and architecture if you work on OS or model safety boundaries.
• Try NucleScript in the browser if compilers or DNA storage are your direction.
• Sponsor physical validation, clean-machine testing, and open hardware research through GitHub Sponsors.

Build ambitious systems. Measure what they actually do. Publish the gap.