ABL-000 · SELF-HOSTED MIND · DESCENT VECTOR LOCKED
Custom agents, workflow automation, and self-hosted LLM deployment — shipped as working software, not slideware. From the model weights up to the browser automation that drives the whole thing.
LOG // 1 CRAFT INBOUND · PILOT: A. HOSSAIN · DESTINATION: LAB
29.9M
parameters pretrained from random initialisation — 2.0B tokens, 9.0 hours, one RTX 3070 Ti. Best val loss 1.341. Own architecture, own tokenizer, own weights.
27B
served on that same 8 GB card — CUDA llama.cpp from source, GGUF quantization, MTP self-speculative decoding. Measured, not guessed →
128
tests reported in the August 2026 snapshot, alongside 7,600+ records and 450+ commits. These historical counts do not establish the current release, deployment gate or runtime.
36/0/0+31
source-snapshot categories built / partial / scaffold, graded by a published rule → — 300+ lines and no stub markers. This measures source presence, not whether an agent is wired, tested, deployed or healthy. The +31 are specified and not built — each with the files it would touch and how it fails safe — counted separately, because a specification is not an agent.
REC 01 // MISSION RECORD
A personal-AI lab coordinating specialized agents through a FastAPI backend, local model routes and a vector-searchable store. Application-level egress controls apply to supported routes; they are not a machine-wide network firewall. Any external provider route needs its own data and authorization checks.
REC 02 // MISSION RECORD
Separated analytics, data, decision, execution, simulation, dashboard, and testing layers. Sole-authored the React 19 + TypeScript operator dashboard, plus SQL-backed ingestion and evaluation pipelines with logging and reporting.
REC 03 // MISSION RECORD
Running 27B–35B models on a single 8 GB consumer GPU: CUDA llama.cpp built from source, GGUF quantization, MTP self-speculative decoding, and MoE expert-offload tuning. Measured, not guessed.
REC 04 // MISSION RECORD
Discord bots on live servers: 80+ registered operations across 20+ modules, persistent per-user state, automated moderation, and an authorization layer with permission tiers that gates destructive operations behind explicit approval.
Git throughout. Readings use a dated snapshot; after 15 minutes, or an invalid timestamp, current state is unknown.
29.9M parameters, my own architecture code and my own tokenizer, random weights to coherent prose in sixteen minutes — nine hours on one RTX 3070 Ti, including the two silent bugs that made the first attempt eighty-six times too slow.
TX 04.1Six experiments on a single RTX 3070 Ti. The write-up reports multi-token prediction increasing throughput from 1.14 to 1.70 tok/s (about 49%), speculative decoding made things worse, and the biggest win was a Windows setting silently moving VRAM into system RAM.
TOUCHDOWN · ALT 0 M — CRAFT SECURED · PILOT HOME
■ PRESSURE EQUALISED · ENTERING
Nobody let me in. Nobody was here. The room holds the last published snapshot; it cannot show what happened while I was away.
Tell me the problem in a sentence or two. If it's a fit I'll come back with a scope — usually same day.
Not here to hire? Everything else — TikTok, Discord, Hugging Face, and who I actually am — is on about & signals.
THIS PAGE · MEASURED IN YOUR BROWSER
Hand-written HTML, CSS and JavaScript — no framework, no build step, no component library. These are read from your browser when the page finishes loading, so they are yours, not mine: reload and they move with your connection and your cache.