Augmenta

Overview

Augmenta is a proof-of-concept privacy layer designed for LLM workflows. It addresses a real gap in production AI systems: the need to handle personally identifiable information (PII) safely before it ever reaches a language model.

Problem

Sending raw user inputs to LLM APIs exposes sensitive data to third-party services. Most teams handle this ad hoc — or not at all. Augmenta explores a structured approach: detect PII, anonymize it deterministically, forward only the clean text to the model, then rehydrate the response before returning it to the client.

Architecture

The system is organized as a set of focused services:

• Ingestion API: Receives raw input, runs PII detection, and orchestrates the anonymization pipeline.
• Anonymizer: Replaces detected entities with stable pseudonyms, maintaining referential integrity across multi-turn conversations.
• LLM Gateway: Forwards anonymized prompts to the target model. The raw PII never reaches this layer.
• Rehydrator: Restores pseudonyms to original values in the model’s response.
• Audit log: Tracks every transformation step for observability and compliance.

Design Decisions

• Structured outputs: All API responses are validated with Pydantic schemas to prevent malformed data from propagating.
• Deterministic pseudonyms: The same entity maps to the same pseudonym within a session, enabling coherent multi-turn rehydration.
• Service-oriented boundaries: Each concern is isolated behind a well-defined interface, making components testable and replaceable independently.
• No data persistence by default: PII is held only in memory for the duration of a request unless explicitly configured otherwise.

Key Learnings

Building this surfaced the tension between LLM flexibility and structured data guarantees — a recurring challenge in production AI systems. Robust parsing of model outputs, handling of partial anonymization, and managing audit scope without over-logging are the hard parts.