Open-Weights Agent

A Next Right Thing proof of concept · two parallel branches · Lucille manages, Roper sponsors · begun July 22, 2026

Can we stand up an agent on an open-weight AI model — with a front end Roper can message from anywhere in the world?

That is the whole project. The test is the same for both branches: if Roper can ping it from Kansas City and get a good answer back, that branch succeeded.

Now: China branch — Step 1. Kai reads the plan and gives his yes or no, then scopes the hardware (buy vs. rent) and writes the short report.

Now: Western branch — success test passed. A password-protected web chat is live at 167.71.202.224, plus a Telegram bridge (@nrt_openweights_bot). Learning write-up delivered July 24. Next: the hourly-GPU mission.


The two branches

One question, two attempts at it. Different models, different countries. Whichever branch lands first teaches the whole team — and neither branch waits on the other.

China branchWestern branch
BuilderKai (Shenzhen)Lucille + Narra (Philippines)
ModelsDeepSeek, Qwen, Kimi, GLMLlama, Gemma, or Mistral
MachineBuy or rent — Kai scopes itDigitalOcean droplet — 4GB, Singapore, $24/month
BegunJuly 22, 2026July 23, 2026
StatusStep 1 — scope the hardwareSuccess test passed — Steps 1–3 complete, hardening shipped

China branch — Kai

The strongest open-weight AI models in the world right now come from Chinese labs — DeepSeek, Qwen, Kimi, GLM. Most US companies cannot or will not touch them. Kai lives inside the Chinese firewall. He has native access to these models, their documentation, and local pricing. For this project, that is an advantage almost nobody else has.

Step 1 — Scope the hardware ← now

Kai answers one question: what machine do we need — buy or rent?

Deliverable: a short report with prices and Kai's recommendation, in about two weeks. Roper reads the report and approves the hardware before any purchase.

Step 2 — Stand up the agent

Get one Chinese open-weight model running on the chosen hardware (Kai picks the model). Make it an agent: it answers questions and can do at least one useful task.

Step 3 — Put a front end on it

Roper can message it from anywhere in the world — phone or laptop — and get answers. Kai picks the simplest path that works across the firewall. Deliverable: Roper pings it from the US, it answers, and Kai writes a short summary — what it cost, what he learned, what he would do next.

Western branch — Lucille + Narra

The Western open-weight family — Llama, Gemma, Mistral — is the other half of the map. Lucille tests it from the Philippines, with Narra as research partner. And simply knowing how to stand up a machine to run an open-weights model is a valuable skill on its own, no matter what happens with anything else.

Step 1 — Learn and choose the path

Done ✓ — compared her own computer against a small rented VM, picked a DigitalOcean droplet (4GB/2vCPU, Singapore, $24/month) and Llama 3.2 3B as the starter model. Roper approved path and spend July 23.

Step 2 — Stand it up

Done ✓ — droplet live, Ollama installed, Llama 3.2 3B answering questions correctly.

Step 3 — Front end reachable worldwide

Done ✓ — and tested. A password-protected Flask web chat page went live at 167.71.202.224. Roper pinged it from Kansas City the same day and got a good answer — branch success test passed. A Telegram bridge (@nrt_openweights_bot) followed right after, so Roper can message the agent like a contact instead of, or alongside, the web page.

Next: Mission 2 — test whether a much larger (70B-class) model can be rented by the hour on a GPU, tested, and torn down the same day. Lucille's learning write-up — what worked, what surprised her, what she would tell Kai — was delivered July 24.

No deadline pressure on either branch — but if a step is stuck for more than two weeks, we pause and regroup instead of grinding.

Roles

PersonOwns
Kai — builder, China branchAll technical choices on his branch: hardware recommendation, model, setup, front end. Friendly advisor to the Western branch when he has time — that branch does not depend on him.
Lucille — builder-learner + managerBuilds the Western branch. Manages the whole project: the plan, check-ins with Kai at least twice a week, keeping this page current, tracking spend and hours, raising blockers to Roper.
Narra — research partnerLucille's co-pilot on the Western branch: research, comparisons, write-ups.
Roper — sponsorDirection, funding, the go/no-go at each branch's Step 1 gate.

Hard rules — both branches

  1. Chinese law comes first (China branch). Private project for us only — no public service, no bot open to users in China. If anything needs registration or ID verification Kai is unsure about, or feels legally risky in any way, he stops and flags it. Kai is never asked to take legal risk.
  2. Absolute data wall. No NFP data, no client data, no personal data — ever. Test data only. Assume anything sent to a hosted Chinese API may be kept.
  3. Budget. Running costs capped at $250/month, shared across both branches. Western branch VM target: under about $40/month. A hardware purchase is separate — one-time, approved by Roper. Lucille tracks all spend.
  4. Security basics. Private repo only; no passwords or API keys ever committed; keys live in a password manager or a local .env file.
  5. Ownership. Everything built on paid NRT time belongs to NRT.
  6. Document in simple English as we go.

Decision log

DateDecision
Jul 23, 2026Western branch: droplet stood up the same day it was approved — Llama 3.2 3B live behind a password-protected web page, Telegram bridge (@nrt_openweights_bot) added. Roper tested from Kansas City — success test passed.
Jul 23, 2026Second branch added. Lucille + Narra will build the Western branch — Llama, Gemma, or Mistral — from the Philippines, on her own computer or a small rented VM. Narra restated the plan; Roper approved. Kai is a friendly advisor to this branch, not a dependency.
Jul 22, 2026Project agreed. Scope shared with Lucille and Kai; Lucille confirmed the plan (via Narra restatement) and accepted the manager role. Budget cap set at $250/month. Check-in rhythm: at least twice a week.

Reports

Report 1, Western branch — own computer vs. VM, with prices. Delivered July 23 (DigitalOcean droplet chosen).
Learning write-up, Western branch — what worked, what surprised Lucille, what she would tell Kai. Delivered July 24.

Report 1, China branch — hardware scoping. Coming after Step 1.