Zero to production: Building a secure AI application without a team or a budget

What happens when a product security professional builds an end-to-end AI security application — alone, on consumer hardware, using AI-assisted coding — and makes it production-ready? This session is the answer. Not a theory. A working pipeline: fine-tuned models ? secure application wrapper ? structured evaluation harness ? regression testing ? deployment. Built on a home grade system. Zero cloud compute. Total rebuild time from a wiped machine: under one month!

The starting point: four AI models — including a 70-billion parameter security reasoning engine — fine-tuned on 229 ISO 27001 and NIST CSF controls, enriched with 1,832 training examples, running locally on Apple's open-source MLX framework at zero cost. These are not general-purpose models prompted to answer security questions. They reason within the framework — mapping relationships across six major frameworks simultaneously through the Control Mesh. The application wrapper — built with Next.js, TypeScript, and Supabase RAG — exposes these capabilities through a product interface with access controls, session management, and structured API responses. Live at vik.so.

What makes this talk valuable is not the demo. It is what the build process reveals about how product security must change when AI is in the stack.

When infrastructure is provisioned from a natural language prompt, and code is generated by an AI assistant, the security professional's role shifts: from reviewing outputs to authoring the constraints the AI operates within. Security requirements embedded into training data, system prompts, and evaluation criteria become the baseline for everything built on top. That is the most consequential design function in an AI-native product organisation.

Four critical decision points every team building AI-assisted security products will face:

1. Training data quality is the security baseline. What enters the fine-tuning dataset determines what the model believes about security. Garbage in, governance failure out. I will show how to structure training data for domain fidelity — and what happens when you don't.
2. The application wrapper is where compliance lives. The model does not enforce access controls, data residency, or session isolation. The application does. I will show the secure wrapper architecture: authentication, rate limiting, structured output validation, and what breaks first under adversarial input.
3. Evaluation is not optional — it is the security control. A model that produces a plausible answer is not a model that produces a correct one. Result-laundering — shipping the right output without verifying the right process — is the failure mode nobody sees coming. I will demonstrate a structured evaluation harness: golden datasets, regression detection, and how to catch silent failures that feel like success until a client's data surfaces them.
4. Vibe-coding has failure modes your SDLC wasn't designed to catch. AI-assisted coding accelerates the build and obscures the assumptions. The teams that survive the production reckoning test the process, not just the output.

​