◢ LOCAL-FIRST // AI + SECURITY

AI and security systems that run on one machine.

I design and build tools end to end — from the inference engine and data model up to the desktop shell. The through-line is systems that run entirely on your own machine: dependency-light, auditable, and tuned to do real work on ordinary hardware.

Focus

systems & tools engineering

Domains

local AI · applied security

Stack

Python · Node · Rust / Tauri

Runs on

one desktop, one GPU

◢

Nirry Labs — live preview

PRIVATE BUILD

nirry-labs // local control serving

GPU arbiter16 GB · holder: language

llama-3.1-8b (active)7.4 GB
desktop reserve1.4 GB
free / handover7.2 GB

Models1 active · 3 idle

llama-3.1-8b-instructQ4_K_Mlive47 t/s
sdxl-turbofp16idle—
ltx-videoQ8idle—
qwen2.5-7bQ5_K_Mqueued—

Throughputtokens/s · 60s

47 t/s

Event logtail -f

Representative interface — the Nirry Labs build is private.

Selected systems

02 / ACTIVE

Local AI platform

Nirry Labs

An all-in-one local AI studio — language, image, video and interpretability — on a single consumer GPU.

Nirry Labs is a self-hosted dashboard that turns one desktop into a complete AI workbench. It serves open-weight language models through a router that hot-swaps models on demand, and layers a full chat workspace, an image studio, short-form video and a multi-model reasoning council on top of the same engine.

The hard constraint shaped the whole design: everything shares one 16 GB GPU alongside a working desktop. So the platform arbitrates the GPU between language, diffusion and video workloads, enforcing clean handover and idle release so no two jobs ever contend for memory.

It ships as plain Node with a buildless web front end — no runtime dependencies and no build step — and is hardened for local use with token-scoped access and keys kept out of the app sandbox.

Zero-dependency engineering. The entire dashboard is vanilla Node and a hand-written web UI, with no framework or bundler to maintain.
GPU scheduling. A single arbiter gives the 16 GB card to inference, diffusion or video in turn, with enforced release and a five-minute idle timeout.
Interpretability tooling. Layer- and neuron-level steering via control vectors, plus a persona library — built as a real research surface, not a demo.
Agentic coding sandbox. Optional Docker/WSL2 environment with live dev-server preview for running coding agents safely against the local models.

Capabilities

Model serving

A local router with per-model presets that hot-swaps GGUF models on demand and streams responses.

Chat workspace

Saved conversations with branch, edit and regenerate, attachments, structured output and approved tool calls.

Image studio

Text-to-image, instruction edits and restyle, each with a per-model pipeline tuned for its quirks.

Short video

Image-seeded clips with synchronized audio, plus chained long-form rendering from a single frame.

Council

A three-stage consensus protocol: independent proposals, anonymous peer review, then a chair synthesis.

Model steering

Control-vector editing at the layer and neuron level, with an explorer for comparing steered outputs.

Security tooling · authorized testing

Pentest Suite

A workflow platform for authorized penetration testing, built so that safety is structural rather than optional.

The Pentest Suite gives a security tester one coherent engagement workflow: discover a scope, enumerate it, correlate what's found against known vulnerabilities, and produce a defensible report. At its core is a headless Python engine backed by SQLite that holds the engagement graph — hosts, services and findings — as a single source of truth.

Every action passes a default-deny scope guard and lands in an audit log, so the tool can only ever touch what the engagement authorizes. Discovery runs on stdlib-only probes (service fingerprinting, DNS, TLS posture, SMB/LDAP/SNMP, web content discovery) and can also orchestrate external scanners, normalizing their output back into the same graph.

Findings are correlated against NVD, CISA-KEV and EPSS and ranked by real-world exploitability, then carried through a lifecycle with retest diffs and printable reports. The whole engine is wrapped in a Tauri desktop shell and frozen to a single binary, so it installs without a Python runtime.

Safety by construction. Scope checks, impact-tier opt-in, per-job approval and governance all converge on one execution choke point enforced before any command runs.
Pluggable execution backend. The same validated command can run locally, in a container or over SSH, without the driver ever being able to widen the guardrails.
Governance controls. A kill switch, testing windows and rate limiting are evaluated ahead of scope, so an engagement can be paused or bounded in time instantly.
Evidence you can stand behind. Every run captures hashed output into per-engagement evidence, and reports carry an exec summary, finding status and a retest diff against a baseline.

Capabilities

Engagement graph

A single SQLite model of hosts, services and findings that every adapter reads from and writes back to.

Scope guard

Default-deny target checking on every action, with a complete audit trail of what was touched and when.

Native enumeration

Stdlib-only probes for services, DNS, TLS posture, SMB/LDAP/SNMP and web discovery — no external install.

Tool orchestration

Drives external scanners through validated templates and normalizes their output into the graph.

CVE correlation

Matches discovered software against vulnerability feeds and ranks by exploitability, not just severity.

Lifecycle & reporting

Finding status, retest diffs against a baseline, and printable HTML or Markdown reports with evidence.

Open source

03 / PUBLIC

Small, self-contained tools that fall out of the work above — each dependency-free, documented, and public. A learning journal that happens to be useful software.

vramlease

Cooperative GPU memory leasing — let local jobs share one GPU by VRAM budget, with crash-safe leases and priority + preemption.

Pythonstdlib-onlyGPU

llm-council

Multi-model consensus: propose → blind peer review → chair synthesis, over any OpenAI-compatible endpoint.

Pythonstdlib-onlyLocal AI

vulntriage

Triage dependency CVEs by real-world exploitability — correlate OSV + CISA KEV + EPSS and rank what to fix first.

Pythonstdlib-onlySecurity

All repositories on GitHub ↗

What this work shows

06 / DOMAINS

Systems architecture

Designing a clean core — a single source of truth, thin interfaces around it — that stays coherent as features stack up.

Local inference & GPU scheduling

Serving open-weight models on consumer hardware, and arbitrating one GPU across competing workloads.

Security engineering

Making safety structural: guardrails enforced at a single choke point, with audit and scope at the center.

Desktop delivery

Shipping cross-platform desktop apps with Tauri and frozen binaries — no runtime for the user to install.

Dependency-minimal builds

Preferring the standard library and buildless front ends, so tools stay small, legible and easy to audit.

Data modeling & correlation

Turning noisy tool output into a normalized graph, then ranking it against real-world signals.