Problem. Voice assistants that call a cloud API for every step aren't private, and stock local models are unreliable at multi-step tool orchestration.

Built. Fine-tuned Llama-3.1-8B via QLoRA on 3,682 examples (1,304 self-authored across 16 categories + a public function-calling dataset), backed by a 54-tool ReAct agent and a 14-gate confirmation layer. Benchmarked against the stock model on 36 held-out workflows: 52.8% vs. 16.7% exact multi-step tool sequencing.

Problem. Turning documents into retrieval-ready knowledge usually means fragile OCR pipelines or opaque VLM extraction with no way to verify what got dropped.

Built. A 7-service multi-tenant production-oriented platform (pgvector, tenant isolation, human approval before indexing) plus a separate research engine with explicit unresolved/ambiguous states instead of silent guesses — 18 extraction engines benchmarked, table reconstruction evaluated at 100% word capture / 96.3–100% row association across 8 documents.

Problem. Native VLM coordinate output for document layout needed to be tested directly before trusting it inside a production pipeline.

Built. 36 versioned geometry experiments comparing native VLM coordinate output against measured page geometry plus adapted model features, with every experiment — including failures — preserved for comparison.

Problem. Tactical-vehicle route-risk analysis needed roughly 10 m terrain resolution, but only 30 m elevation data was available.

Built. Trained a 15.6M-parameter RCAN with physics-aware loss — rejected adversarial/GAN-style super-resolution to avoid hallucinated terrain detail that would corrupt downstream slope/curvature calculations — across 900 tiles / 5 US regions. Colorado test region: −43.8% elevation RMSE, −69.9% slope RMSE, −86.2% curvature RMSE vs. baseline.

Problem. Full AI autonomy over appointment scheduling breaks down at ambiguous intent, stale availability, and irreversible changes — those need deterministic controls or a human, not a model's judgment alone.

Built. A bounded AI agent with 7 typed tools validated by a state guard, a 14-state workflow/lifecycle engine, idempotent BullMQ job scheduling, and staff-in-the-loop controls treating "the AI should stop" as a first-class outcome.

Problem. Small local models struggle with tool-calling as a tool registry grows — both calling a tool nobody asked for, and picking the wrong one among similar options.

Built. An embedding-based semantic tool router (top-3 of 19 tools by cosine similarity), a Phase-1 ACTION/QUESTION intent classifier, and disambiguation-by-prompt across near-identical tools — running inside a daily-use Photo/PDF/Video editing suite.

Problem. Local coding models need controlled access to files and shell operations rather than unrestricted execution.

Built. Terminal-native Python agent with 9 structured tools, 5 mode-scoped permission registries (read-only modes remove write/shell tool schemas from the model entirely, not just by prompt), atomic writes with timestamped backups, and a regex-based XML fallback parser recovering malformed Qwen3-Coder tool calls.

Problem. Capturing a live site's real DOM, network activity, and source maps for offline analysis needs more than a static scrape.

Built. A CLI that drives a real browser session over CDP to capture DOM state, network traffic, shadow DOM, scripts, source maps, and visual evidence. The archive is designed for offline analysis; API/XHR responses are recorded but are not fully stubbed by the replay server.

Problem. Job searching and resume tailoring usually means sending personal career data to a cloud service.

Built. A local-first job search and resume-tailoring tool built on Flask and Ollama, keeping resume data and model inference on-machine while exposing explicit diagnostics for its external job-board sources.

Problem. Tracking job applications and judging fit against a resume is manual, repetitive work.

Built. A job tracker with AI-assisted fit analysis, local JSON persistence, local Ollama model access, and a companion browser extension for capturing listings from logged-in pages.

Problem. AI coding agents are useful and unreliable in specific, predictable ways — each of these five systems closes one of those gaps rather than trusting the agent blindly.

• Memory System — persistent project state, decisions, and failures so an agent reads indexes before full files. In use across 20 of my own workspaces. MIT licensed.
• Security Testing OS — a gated, question-driven workflow (intake → threat model → architecture review → release gate) run before and during a build, not as a final scan; explicitly tracks what's not yet built.
• Software Planning OS — risk-tiered planning that produces validation artifacts, not description artifacts, to catch scope creep before code.
• CodeArch — reverse-engineers any codebase into structured documentation; every claim is sourced to a specific file or explicitly marked "not found," never invented.
• Idela — an idea-to-architecture workflow synthesizing named practices (Amazon's Working Backwards, Google Design Docs, Arc42, C4, MADR, RFC process); still early-stage in real-world use.

Problem. Choosing a document/OCR/layout model on reputation instead of measurement risks building on the wrong foundation.

Built. Isolated benchmark workspaces comparing Nemotron, PaddleOCR-VL, NuMarkdown, Ministral, and layout detectors. Each track records its own inputs, settings, outputs, and failures; aggregate comparisons should be read with those track boundaries in mind.

Problem. Standard screenshot tools don't offer free-form region selection for quickly capturing exactly what's needed for a bug report or AI prompt.

Built. A Linux capture daemon with fullscreen, window/monitor, rectangle, and free-form lasso selection, immediate clipboard copy, auto-save, and frozen frame capture before the picker appears. Verification is currently manual, with focused geometry unit tests.