Privacy Program Manager by profession. Local-AI researcher by obsession.

My career protecting user privacy taught me something uncomfortable: the AI tools we rely on daily are architecturally incompatible with meaningful data sovereignty. Every query, every document, every personal thought fed into a cloud assistant traverses infrastructure you don't control, governed by policies you didn't write.

So I'm building the alternative.

BlarAI is my primary research project: a security-first, multi-agent AI system designed to run entirely on a single consumer laptop — no cloud, no telemetry, and zero external network dependency by default. When optional web search is enabled, every outbound call goes through one audited, deny-by-default egress door.

▶ Watch it run (2 minutes):

Signature-verified boot → local chat with live tool calls → document ingestion → vision → on-device image generation → voice in and out → a fingerprint-gated web door → a refused destructive request → BlarAI's own coder fleet building a real website. Every image and voice in the film was generated by BlarAI, and the score was synthesized on the same laptop. (public repo · full-quality stream · 1080p download)

▶ New — the vision film (66 seconds):

The system runs on an Intel Core Ultra 7 258V (Lunar Lake) with 32 GB LPDDR5X and an Arc 140V GPU, and includes:

• Policy Agent — A hybrid deterministic + probabilistic security gatekeeper that intercepts and classifies every inter-agent action before execution. Fail-closed by default.
• Private Assistant — A conversational AI with a streaming WinUI desktop app, voice in and out, modular skill routing, and a four-layer defense-in-depth against indirect prompt injection.
• Personal Knowledge Substrate — Semantically indexed local knowledge base with constant-time retrieval pacing to suppress timing side-channels.
• Headless Coding-Agent Dispatch — Decomposes a natural-language goal into tasks with verifiable acceptance criteria and dispatches them to a local coding fleet. Local subprocess only, zero egress.

The architecture enforces hardware-rooted trust (a TPM 2.0 trust root with signature-verified boot), cryptographic action authorization (single-use action tokens with a 128-bit nonce and a 5-second time-to-live), and mutual Transport Layer Security on every inter-service channel — layered so that no single-layer failure can authorize a malicious action.

Built with OpenVINO · OpenVINO GenAI · Python · Qwen3 · Intel Arc GPU · Intel NPU

I contribute upstream to the tools I depend on — fixes, testing, and measurement, always from real workloads running real models on real hardware. Status shown for every item, because most of what I submit does not land and the record is more useful with that left in.

Merged

• openvinotoolkit/openvino.genai#4082 — PR (merged, July 2026): fix an xgrammar structured-output crash at EOS under speculative decoding. Verified from 1 crash in 95 generations to 0.

Open

• openvinotoolkit/npu_compiler#302 — PR (open): unique per-consumer slice locations in UnrollGroupQuantize — the root-cause fix behind #265/#266 below
• openvinotoolkit/openvino#34617 — Filed: compile_model fails with dynamic shape error for Qwen3-0.6B INT4
• openvinotoolkit/openvino.genai#4091 — Filed: feature request to revive persistent, disk-backed KV-cache dump/restore for cross-session prefix reuse, with independent iGPU benchmarking offered

Closed

• openvinotoolkit/openvino#34651 — PR (closed unmerged, July 2026): early guard for unbounded dynamic shapes in NPU compile_model, proposed against #34617
• npu_compiler#265 / #266 — PRs: zero-dim guards in ConvertFCToConv / UnrollFullyConnected (closed by me, June 2026 — a maintainer pointed out these guarded where the crash surfaced rather than where the bad tensor was created, so I withdrew them for #302)
• openvinotoolkit/openvino#34450 — Filed: LLVM abort in as_convolution pass for Qwen3-0.6B INT4 on Lunar Lake NPU
• openvinotoolkit/openvino.genai#3429 — Filed: LLMPipeline NPU draft model abort on Lunar Lake

I also publish measurements as open data — two CC-BY-4.0 datasets on HuggingFace characterising local inference on Lunar Lake, including a negative result showing the vendor-documented NPU draft path is slower than not using the NPU at all.

Every bug I file comes from real workloads running real models on real hardware — not synthetic benchmarks.

I practice AI-assisted development — I architect systems, make design decisions, and direct AI agents to implement, test, and operationalize the software. The current toolchain: Claude Code (Anthropic) for interactive engineering sessions, a fully-local coding fleet — OpenCode driving a 30B coder model on the Arc 140V (agentic-setup) — for autonomous overnight runs, and GitHub Copilot for review passes. Every line of code in BlarAI was produced through human-directed AI collaboration: I own the architecture and every technical decision; AI agents own the keystrokes.

I'm transparent about this because I think it's the future of software engineering — and because the results speak for themselves: a standing test gate of 10,000+ tests at zero failures, formally specified use cases, every claim hardware-validated on real silicon.

My passion for privacy fuels my research with local AI. I believe the average consumer device — the laptop already on your desk — can run powerful AI tools privately and securely, without surrendering your data to someone else's server.

I seek to contribute to open source projects that make this a reality for everyone.