ModelBrew provides a continual learning platform that lets users fine‑tune a single base model with new data in real time, without forgetting previously learned information. The system stores each customer’s data in isolated adapters, enabling instant, provable erasure that generates cryptographic certificates and a tamper‑evident audit trail to meet GDPR and emerging AI data‑deletion regulations.
Funding
Funding not disclosed
Founders
Product
Problem
Fine‑tuning large language models typically leads to catastrophic forgetting, requiring costly retraining to incorporate new data. Additionally, emerging data‑deletion regulations (e.g., CCPA, GDPR) demand provable erasure of personal information embedded in model weights, which traditional fine‑tuning cannot guarantee.
Solution
ModelBrew offers a continual‑learning AI platform that enables users to add new facts to a frozen base model in a single live pass without retraining, preserving existing knowledge. The system isolates each tenant’s data in a self‑contained adapter, so removing the adapter instantly unlearns that information. Every deletion generates a recomputable cryptographic certificate and is recorded in a hash‑chained audit log, providing verifiable proof of erasure and tamper‑evidence for compliance. The platform also includes built‑in security controls such as role‑based access, modern authentication, and encrypted backups, allowing AI teams to manage fine‑tuning, preference tuning, and continual learning without managing infrastructure.
Target Audience
Primary customers are AI development teams in regulated sectors such as healthcare, finance, legal, and enterprise SaaS that need to continuously update LLMs while meeting strict data‑deletion and audit requirements.
Features
- Single‑pass live continual learning that updates model knowledge without catastrophic forgetting
- Adapter‑based architecture that isolates tenant data on a frozen base model for instant, provable unlearning
- Cryptographic deletion certificates (SHA‑256) that can be independently recomputed to verify record erasure
- Hash‑chained audit trail logging all privileged actions for tamper‑evident compliance reporting
- Integrated dataset cleaning with 60+ validators, AI‑judge rewrites, and PII‑heuristic filters
- Self‑serve fine‑tuning (LoRA/QLoRA) and preference tuning (DPO/SimPO) on multiple 7‑9B open‑source models
- Serverless GPU execution with flat‑rate pricing ($3.99 per million tokens) and OpenAI‑compatible endpoints
- Enterprise‑grade security: bcrypt‑hashed passwords, optional TOTP 2FA, per‑user/IP lockout, role‑based API keys, encrypted backups, and VPC/on‑prem deployment options