FR-Start / README.md
FahrenheitResearch's picture
Upload README.md with huggingface_hub
22a8139 verified
|
Raw
History Blame Contribute Delete
2.25 kB
metadata
license: gemma
base_model: google/gemma-3-12b-it
language:
  - en
tags:
  - legal
  - tax
  - incorporation
  - startup
  - ollama
  - gemma3

FR-Start — Startup Incorporation Advisor

Fahrenheit Research · Runs 100% locally via Ollama · No API, no cloud, no data leaves your machine.

FR-Start tells startup founders where to incorporate and why, across the United States, India, the UAE, Singapore, and the United Kingdom. It runs a structured intake, then applies a fixed decision priority — investors > market access > founder tax residency > compliance cost > tax — grounded in a curated, dated knowledge corpus covering:

  • corporate & indirect tax, entity types with pros/cons
  • compliance load, setup cost/time, banking friction
  • hiring & payroll costs, founder personal tax, exit & M&A treatment
  • grants & incentives, cross-border rules (flips, PE/POEM, withholding, transfer pricing, FEMA)
  • step-by-step incorporation playbooks per jurisdiction

Architecture

This is a knowledge-augmented model, not a fine-tune: the base model is Gemma 3 12B (12 billion parameters), and all regulatory facts live in an auditable Markdown corpus fused into the model's system layer at build time. Facts carry as_of: 2026-08 dates (web-verified August 2026). Updating a tax rate is a one-line corpus edit + rebuild — no training run.

Use

ollama pull gemma3:12b
ollama create fr-start -f Modelfile
ollama run fr-start "SaaS founder in Bangalore raising from US VCs — where do I incorporate?"

Needs ~10 GB RAM. The corpus/ directory in this repo is the full knowledge base — edit it, bump the as_of dates, and re-run ollama create to refresh.

Limitations

  • Not legal or tax advice. Rates and rules change; every answer carries as-of dates and must be verified with a qualified professional before acting.
  • Covers only the five listed jurisdictions.
  • A 12B local model: reliable on corpus-grounded questions, thinner on novel multi-jurisdiction edge cases than frontier models.

Base model weights are Google's Gemma 3, used under the Gemma Terms of Use. This repo distributes only the knowledge corpus, Modelfile, and tooling.