Adriel's Lab > Roadmap

The current arc: turning ~30 working projects into one engineered system — a control loop that reads declared intent, routes work to the right trust lane, and verifies outcomes in a domain the worker can’t touch.

1. Make the fleet boringly recoverable. Everything in version control with a contract, or explicitly out of scope. Done — every active project repo has a HEAD and a contract.
2. Prove the money lane. Run the tax engine against real exports and tie the result out against a commercial tool. Done — independently confirmed.
3. One complete control loop. One project, one worker, one acceptance signal — plus a full failure-state machine, because autonomy without defined failure states is just an unhandled exception with write access. Paused — waiting on the AI server.
4. Instrument immediately. Canonical event log, metrics as rebuildable projections. Paused — waiting on the AI server.
5. Generalize, then demonstrate. The loop becomes the instrument that measures the workers — and the write-up becomes the proof. Paused — waiting on the AI server.

Machines

The hardware side of the same roadmap. Deliberately a task list and not a schedule — some of these need money, some need a free weekend, and I am not going to pretend I know which month either shows up in. What the fleet runs on today is on the Homelab page.

• Build the EPYC AI server. The plan is locked: an always-on platform with three GPUs and 60GB of VRAM, priced and reasoned out against the alternatives before anything was bought. Parts get acquired as they get affordable. Not built — the build plan is done, the machine isn’t.
• Fix routing before buying capacity. The 24GB card is idle far more often than it is saturated, so the real bottleneck is scheduling work onto the right machine, not a lack of machines. Buying a second GPU to solve a routing problem would just produce two idle GPUs. Open — and it comes before the shopping.
• Give the GPU a real queue. Interactive chat, transcription and corpus grading all contend for one card with nothing arbitrating between them. Batch work should yield and interactive work should preempt. Open.
• Schedule across the whole house. The end state: the orchestrator places a job on whatever node can run it and happens to be awake, instead of everything landing on the workstation. Idea — needs the queue above to exist first.
• Shrink the OptiPlex boot drive. It runs a 1TB disk to hold what should fit on a small SSD, but 293GB of service data has to move to the array before that swap is possible. Blocked on the migration, not on the hardware.
• Put the borrowed rigs to work properly. Two family gaming PCs are the only real burst capacity here, and both are intermittent by nature — they are someone else’s computers first. Opportunistic by design; will never be dependable, and shouldn’t be planned as though it is.

Newspapers

Extracted information into the corpus every week or month.

1. Bitcoin
2. AI
3. Mindset
4. Real Estate Doomsday Dossier
5. Parody defi investor agency

Straightforward Workflow:

1. Local LLMS do most of the grunt work: find sources, find the signal, and put them together based on the instructions.
2. Cloud models are the editor and verifier. Does the finishing touches and verifies information and claims. Like Grok, or Claude.

Corpus:

The database of reference for everything.

Factors to consider:

1. Grading for quality
   • Creator
   • Video/podcast
   • Short term information, or long term
   • Verify claims