Skip to content
BASANOS
Tier 03 · Advanced

Own the model that runs your workflow.

RL fine-tuning of open-weights models — Qwen-class, GRPO — in a sandboxed clone of your actual workflow, with your hardened judge as the reward.

The loop

An eval the model is allowed to practice against.

Every episode: the policy reads an account, proposes a plan, drafts the reply, gets scored, and nudges its weights toward what scored well — ten thousand times. Reward-hacking attempts show up, and die against the judge hardened in Tier 02.

The reinforcement-learning loopA Qwen-class open-weights policy acts inside a sandboxed replica of the servicing stack. Each episode transcript is scored by the hardened judge as the reward; a GRPO update shifts the weights toward high-reward behavior, looping ten thousand episodes. Result: task success rises from 61 to 88 percent against the frontier-API baseline, at roughly ten times lower cost per ticket, with the client owning the weights in their own VPC.QWEN-32B POLICYHarper-in-training · open weightsactsSANDBOX ENVIRONMENTmock CRM · policy DB · simulated borrowersreplica of the real servicing stackepisode transcriptREWARD = HARDENED JUDGEscoreGRPO UPDATEweights shift toward high-reward behavior× 10,000 episodestask success 61% → 88% (vs frontier API)~10× lower cost per ticketthe client owns the weights, in-VPCverified on the Tier 01 harness

Plain English, for the people who sign

GRPO
The training method. What matters to you isn't the acronym — it's the grader it trains against.
Open weights
A strong open-source model you can run and own outright — no per-token API bill, no vendor holding your capability.
SFT
Supervised fine-tuning: teaching the model from vetted examples of the job done right, before RL sharpens it.
Reward model
The judge, given consequences: it scores every practice attempt, and the model's weights move toward what scores well.

The sandbox

Your workflow, cloned. Your data, staying put.

We replicate the stack the agent actually lives in — mock CRM, policy database, simulated customers with messy, realistic accounts — so the model trains on the job it will do, not a benchmark that resembles it.

Training runs inside your VPC. Nothing leaves: not the traces, not the policy corpus, and not the weights — because you own them.

Prerequisites

You can’t buy rung three without rungs one and two — the loop is built from their deliverables.

  • In place: Scored trace corpus (Tier 01 — the harness)
  • In place: Hardened, reward-grade judge (Tier 02 — the meta-eval)
  • Next: Training run — GRPO against the judge, in the sandbox

Results

Composite case study — illustrative

Frontier API baseline

61%

Qwen-32B after GRPO · hardened reward

88%

+27pts

Task success vs frontier API

~10×

Lower cost per ticket

Weights: yours, in-VPC

Certified on the Tier 01 harness

This is a real services category

Cursor post-trains models with its own product as the environment; services shops run this exact loop for enterprises — typically on Qwen-class models — for workflows like CRM operations. Most organizations have no in-house RL expertise, so they buy it.

  • RunRL
  • Osmosis
  • Applied Compute
  • Adaptive ML
  • Cursor (self-environment post-training)

Landscape — not clients or endorsements

Start at the stone — book the audit.

Rung three is built from rungs one and two. Every engagement starts with a two-week eval audit.

Book an eval audit

hello@basanos.ai · Replies within 48h