Complete Project Repository

A hardened, unprivileged container environment for local open-weight models whose weights were never fully trusted. The inference server runs with zero Linux capabilities, a read-only model store, and no route off the local bridge — the host reaches the API at a static container IP, with no published ports at all. A twelve-point audit gate verifies the live container against the policy before any request is routed.

The auxiliary engineering around the stack's AnythingLLM workspace: custom model registration, generated-file management, and push notifications — the pieces most deployments take for granted and few actually check.

A Google-style search interface over a self-hosted RAG stack. Users type a plain-English question and get an answer with cited sources — drawn entirely from documents that live on the local stack.

A field log of running self-hosted AI on hardware that fights back: a 2013-era Xeon without AVX2, CPU-only inference, CGNAT, and broken IPv6. Every entry records the symptom, the investigation, and the fix.

A slice of the lab server's operations expressed as code, so it can be rebuilt from a declaration instead of by hand. The Terraform module uses the Docker provider to manage a small, additive monitoring footprint while leaving the existing compose-managed stack untouched.

A Docker Compose stack that provides the services behind most of this portfolio: a vector database (Qdrant), a graph-memory service, a private search engine, and a media server — all sitting behind an encrypted mesh network.

A self-hosted web crawling and scraping stack with API and CLI access, running behind the encrypted mesh. It serves as research ingestion — turning web content into clean, structured data that the rest of the pipelines can consume.

The escalation tier of the handwriting pipeline: when the primary machine's OCR engines can't agree, the disputed pages are shipped to a heavier model on a second machine over the encrypted mesh, and the results are merged back into the primary output.

Deployment and integration of a graph-backed agent memory service into the local stack. It gives the system persistence that's structured as a graph rather than just a vector store — relationships between entities are first-class, so memory isn't just "relevant text" but "who is this and how do they relate."

A multi-engine OCR system for handwritten documents. It runs several open-source recognition models in parallel — PaddleOCR, Surya, EasyOCR, and TrOCR — and merges their results through a consensus layer instead of trusting any single engine.

A dual-layer, fully local memory system that gives the AI stack persistent context without sending anything to the cloud. It combines spatial memory (vector storage with ChromaDB) with an execution-time RAG filesystem (viking://) organized into tiers.

An application of Recursive Language Model (RLM) scaffolding to handwriting correction. Instead of correcting each page in isolation, the whole document's OCR output is loaded into a REPL-style environment and processed as one coherent body of text.

A small shared memory layer for the whole local AI stack: one Qdrant store, local embeddings from Ollama, and a simple CLI (plus an MCP server surface) so any pipeline or agent can remember and recall.

An end-to-end automation pipeline for research discovery and outreach: scheduled scans across regional and local sources, LLM-assisted qualification and de-duplication, personalized draft generation, calendar sync, and a curated dashboard of results.

Local, CPU-only LLM inference for the AI stack — no GPU, no cloud APIs. Custom model configurations handle different workloads on the same commodity hardware.

An automated remote job application pipeline: it aggregates live postings from eight legitimate job boards, filters to roles that are remote, US-workable, and $60k+, and builds a tailored application kit per role — including the real form questions pre-drafted from the employer's own Greenhouse listing.

A scheduled scanner and classifier that watches configured sources, extracts structured listings, normalizes their metadata, de-duplicates them, and writes the results to a structured store — all as a self-contained Docker service.

Self-hosted n8n workflows orchestrating the lab's data pipelines, lead generation, and maintenance automation — all running behind the encrypted mesh with no public exposure. Includes hardened integration with a local LLM backend: generation caps and bounded fan-out so a single-threaded model server is never overwhelmed by a parallel batch.

A self-hosted recipe collector. Shared social media recipe links (Instagram, TikTok, Facebook Reels, YouTube Shorts) are watched from an inbox, cleaned into structured recipes by a local LLM, scrubbed of personal data, and published as a searchable section here. Search by the ingredients you have on hand or filter by category.

Ran a systematic over-engineering audit across several Python projects. Deduplicated shared helpers, removed dead stub functions, consolidated repeated error-handling patterns into a single decorator, and dropped an unused dependency. Net result: ~200 fewer lines, same behavior, verified by clean imports and a full pipeline dry-run.