Coming soon•Full & QLoRA Fine-tuning

Specialized Intelligence for High-Stakes Domains.

We research efficient Small Language Models, build autonomous agent architectures, and provide enterprise infrastructure to fine-tune custom models on proprietary data.

Explore Neotropy Forge Read our Research

neotropy · live consoleepoch 3/4

fine-tune · nt-3b-legal · QLoRA 4-bithover the curve

train val

agent · claims-adjuster · graph v12running

hover a node to inspect it

01 — The Three Pillars

Research, training, and agents — one stack

Everything we ship comes from the same lineage: research informs the trainer, the trainer feeds the agents, the agents surface the next research question.

Neotropy AI Lab

Visual builder for autonomous multi-agent workflows with custom behavioral constraints and API deployments.

no PII exportcite every claimhuman-in-the-loop

Explore AI Lab

Neotropy AI Forge

Enterprise-grade Full Parameter & QLoRA fine-tuning on custom datasets for open-weight models.

supported families

LlamaQwenMistralGemma

run #128 · QLoRA 4-bit68%

Full ParameterQLoRAprivate VPCaudit logs

Explore AI Forge

Frontier SLM Research

Novel research on domain-adaptation, distillation, and dataset curation for small, high-efficiency models.

Read the research

Domain adaptation

Curricula that teach small models your corpus 2.3x faster.

Distillation

Committee-verified teachers compress 70B quality into 3B.

Dataset curation

Entropy-scored pipelines that filter noise before training.

02 — Capability Showcase

Watch the pipeline work

Flip between the two products and follow a real job end to end: data in, deployed intelligence out.

Dataset upload

contracts_2026.jsonl 412 MB

clinical_notes.parquet68%

drop corpus files

Quantization

base model

Qwen2.5-7BLlama-3.1-8BMistral-7BGemma-3-4B

method

QLoRA · 4-bitFull · FP16

VRAM before58 GB

VRAM after11 GB

Training logs

step 0120 · loss 1.842 · lr 2.0e-4

step 0240 · loss 1.213 · grad_norm 0.84

step 0360 · loss 0.941 · eval 1.12

step 0480 · loss 0.788 · eval 0.95 ✓ best

step 0520 · ▍

03 — Research

Publications in the pipeline

We publish what we ship. The first papers are in internal review and will land on the research page at launch.

First papers coming soon

in review

QLoRA recipes, distillation results, and agent-safety notes are being written up now — subscribe to the dispatch in the footer to get them first.

Visit the research page

Launching soon. Be first in line.

Forge and Lab are in final hardening. Join the waitlist — the first cohort trains with our research team on launch day.

SOC 2 Type II on-prem available data never leaves your VPC