Learning Loop / ML Systems Engineer

Company Description

VulnautAI is an AI-powered smart contract security auditing platform built for the world's leading Web3 audit firms. We're not building a generic code scanner — we're building a system that gets smarter with every audit, accumulating the institutional knowledge of each firm over time and surfacing it at the moment it's most useful.

We have ongoing betas from top-tier audit firms, a proprietary dataset of 180,000+ vulnerability findings, and a static analysis pipeline in active development. We're approaching revenue and moving fast.

Role Description

You will own the architecture and implementation of our continuous learning system — the mechanism by which auditor corrections, false positive labels, and confirmed findings flow back into the model and make it measurably better over time.

You'll also own the retrieval architecture that powers our historical audit intelligence feature: when a new contract comes in, the system identifies syntactically and semantically similar code from a firm's prior audits and surfaces whether known bugs from those audits are relevant to the current one. This is a founding-team-level role.

What We're Looking For

Learning loop architecture — Design and implement the full pipeline: auditor correction signals → feedback ingestion → fine-tuning or embedding update logic → evaluation harness → versioned model rollout, all within per-client isolated infrastructure

Per-client isolation — Architect the system so each firm's corrections and historical data improve only their own model instance, with no cross-client data contamination (Postgres + Pinecone per-tenant)

Code similarity retrieval — Build the semantic code search layer that matches incoming contract code against a firm's historical audit corpus and retrieves associated findings as candidate vulnerabilities

Evaluation harness — Define and instrument the metrics that confirm the system is actually improving: precision, recall, false positive rate, retrieval relevance, and correction signal quality

Dataset integration — Incorporate our existing 180K+ finding dataset as the foundational training scaffold, working with existing taxonomy and field structure

Qualifications

Demonstrated experience architecting a continuous learning or RLHF-style feedback loop end-to-end — not just operating within one

Production RAG pipeline experience: ingestion, chunking strategy, embedding validation, retrieval relevance tuning, vector database management

Strong Python and ML engineering fundamentals (PyTorch or equivalent, MLflow or similar experiment tracking)

Experience with per-tenant or multi-tenant ML infrastructure — you understand the data isolation requirements and have solved them before

Experience with code embedding models (CodeBERT, StarEncoder, or similar) or AST-based chunking LangGraph or multi-agent pipeline experience Fine-tuning experience on domain-specific code datasets