What the lab actually builds.

What it is. A personal AI system that runs entirely on my own hardware — a FastAPI backend that coordinates a fleet of locally served language models behind one API, with distinct agent personas, persistent memory, and a hard privacy boundary.

How it works. Requests route across model lanes by task: a fast lane for everyday queries, a heavier lane for analysis, each a locally quantized model served through llama.cpp. Agents share a vector-searchable store (7,600+ records) I schema-designed, plus a markdown vault that acts as long-term conversation memory with TF-IDF recall. An egress boundary classifies every outbound request and pins sensitive data to the machine — the cloud is absent by design.

Why it's hard. All of this fits on one consumer GPU with 8 GB of VRAM. Model selection, quantization, offload strategy, and memory pressure are engineering constraints on every feature.

What it is. A futures-market analysis platform split into strictly separated layers — analytics, data ingestion, decision logic, execution, simulation, dashboard, and testing — about 9,300 lines across 43 modules.

How it works. SQL-backed ingestion pipelines pull and evaluate market data on schedule; the decision layer produces structured signals with full logging so every output can be audited after the fact. I sole-authored the operator dashboard in React 19 + TypeScript — live signal views, conviction boards, and volatility screens, now served behind an authenticated edge (Cloudflare Access + tunnel, no exposed ports).

What it taught me. Honest evaluation infrastructure matters more than any single strategy: simulation and reporting layers exist precisely to prove when something does not work.

What it is. The infrastructure practice under everything else: running 27B–35B parameter models on a single RTX 3070 Ti, measured honestly.

How it works. llama.cpp built from source with CUDA for the exact GPU architecture; GGUF quantization; multi-token-prediction self-speculative decoding (a ~50% throughput gain for one flag); mixture-of-experts expert-offload tuning; and the driver-level forensics — like catching Windows silently spilling VRAM into system RAM — written up with numbers in the 27B-on-8GB post.

Current state. A benchmarked model ladder from a 9B dense model at 77 tok/s fully on-GPU to a 35B MoE at ~22 tok/s with experts on CPU.

What it is. Node.js bots on live servers with real users: 80+ registered operations across 20+ modules.

How it works. Persistent per-user state with crash-safe shutdown flushing; automated moderation; an authorization layer with permission tiers that gates destructive operations behind explicit approval; and the same markdown-vault memory layer as ABL, so the bots remember conversations across restarts. One of them speaks — a local text-to-speech pipeline with a custom blended voice and automatic fallback.

What it is. A coaching tool that watches gameplay clips and produces round-by-round feedback.

How it works. Frame extraction and filtering first — death screens and spectating segments are detected and discarded so the model only analyzes rounds the player actually played. A vision-language model then reads crosshair placement, positioning, and engagement decisions, and returns structured coaching notes.

What it is. A Playwright-driven pipeline that handles long, structured, multi-step form workflows end to end — with a verified-facts layer that guarantees it never invents data.

How it works. Every field an automated run can fill is backed by a single source-of-truth facts file; anything not present is left for human review instead of guessed. Runs produce receipts — screenshots and structured logs per step — so every submission is auditable afterward.