SOPHIA XT
Now in private beta

SOPHIACustom

A custom-trained AI for your domain in 60 seconds.

Bring your knowledge corpus. We crystallize it into our LTMi-XT format and train a customer-specific LoRA adapter using anchor-token masking — a training-objective modification we developed and validated to produce a 1.67× OOD generalization advantage over standard LoRA fine-tuning. Combined with forced-anchor decoding at inference, retrieved facts in your domain are preserved verbatim. Deterministic identifier accuracy. Measurable per-customer quality.

Validated methodology

Empirically grounded.

Every claim below is from the published technical report, with all weights, code, corpora, and 504+ per-query outputs released under Apache 2.0. Read the paper →

1.67× OOD generalization

Anchor-token masking produces a measurable, statistically tight out-of-distribution generalization advantage over standard LoRA fine-tuning. Your customer-specific model is better at extending content from your domain.

30-100 second fine-tunes

Optimal training regime is short fine-tunes on small corpora — exactly the per-customer use case. A fine-tune costs us pennies of compute; we pass margin to you.

V/O-only LoRA — half the parameters

Causally validated to outperform full LoRA at half the trainable parameters. Smaller adapters, lower memory pressure, better generalization, identical training time.

Deterministic identifier preservation

Combined with forced-anchor decoding at inference, retrieved facts (part numbers, citations, drug names, dates) are preserved verbatim. Cannot be paraphrased or hallucinated.

How it works

Four steps. End-to-end.

01

Upload your corpus

PDFs, Markdown, structured exports, knowledge bases. Up to 100 MB per ingest.

02

Crystallize to LTMi-XT

Your prose split into self-contained atomic loci. Hierarchical breadcrumbs, byte-offset provenance, deterministic lattice indexing.

03

V/O-only LoRA training

Anchor-token masking applied to your corpus. 100-300 steps. Validated against held-out subset with comparison to standard LoRA baseline.

04

Hosted inference endpoint

Your fine-tuned model deployed at sophiaxt.com/api/v1/customer/{id}. Forced-anchor decoding for deterministic identifier preservation.

Verticals

Where deterministic identifier accuracy matters.

Domains where a paraphrased part number, drug name, or case citation is a real cost — not a minor embarrassment.

Legal documentation

Case citations, statute references, contract clause numbers preserved verbatim

Pharmaceutical / medical

Drug names, ICD codes, NDC numbers, dosage strings — deterministic preservation

Industrial parts catalogs

SKU numbers, model identifiers, technical specs across heterogeneous documentation

Technical documentation

API references, configuration values, version-specific behavior

Internal knowledge bases

Org-specific terminology, policy references, structured procedures

Compliance & regulatory

Citations preserved verbatim for audit trails — regulatory references cannot be paraphrased

Pricing

Setup fee + per-call inference.

One-time setup covers the human-in-loop crystallization quality check + initial training. Inference is priced per-call, no monthly minimums on Hobby and Pro tiers.

Hobby

$99

one-time setup

+ $0.005 / 1K calls

Solo developers, side projects, prototypes

  • Up to 500 loci per corpus
  • 1 fine-tuned model
  • Hosted inference endpoint
  • Basic email support

Pro

$299

one-time setup

+ $0.005 / 1K calls

Small teams, production use

  • Up to 2,000 loci per corpus
  • Up to 5 fine-tuned models
  • Dedicated hosted endpoint
  • Priority email support
  • Validation report per fine-tune

Scale

$999

one-time setup

Custom inference pricing

High-volume programmatic use

  • Unlimited corpus size
  • Unlimited fine-tuned models
  • Multi-region inference
  • Dedicated success manager
  • Custom evaluation suite

Enterprise

Contact

—

—

On-prem, SOC 2, custom integrations

  • On-prem or private cloud deployment
  • SOC 2 compliance support
  • Custom integrations
  • SLA + dedicated infrastructure
  • Source-available licensing options

Private beta

Join the waitlist.

Currently onboarding a small cohort of design-partner customers. Your corpus, your model, your measurable advantage.

Methodology and validation released under Apache 2.0 — every claim reproducible from public code.