Dylan McGuinness

The best finding so far is broken access control on a major CRM: a low-privilege portal member could read every contact record, including email, phone, job title and company, plus the subject and body of logged emails, while the vendor's own modern surfaces correctly told that same user it could read nothing. Three independent controls confirm it, including a positive control that moves the modern surfaces from zero to one while the legacy routes keep returning everything.

Getting there needs an accurate map of what the application can actually do, so the agent recovers the API contract from the code the client is handed rather than from crawling, which surfaces routes and request shapes a crawler never reaches. Then it works that surface as different identities and compares what each one can get.

• Every finding is verified by direct request before it is written. Nothing is inferred from a scanner, and severity is argued down rather than up.
• The discipline is mostly in what gets thrown away. One bulk pass produced 994 findings that were all false positives. Catching that mattered more than any of them.
• One report was drafted as medium trending high, then killed by a positive control against my own account and downgraded to low. Another was sound but never filed, because the bug class was out of scope for that program.

github.com/dylanzonix/ambit

Clay and Apollo hand you a list and leave you to filter, clean and enrich it across several tools. This was one step: say what you want, agents research each row, a verified dataset comes back. I owned the entire backend and the agent system.

• Verification gated billing. If we could not confirm a real website, a live email or a working phone, the row did not ship and the customer was not charged for the attempt.
• Source allocation is a multi-armed bandit. Crawl budget goes to sources by Thompson sampling on per-source Beta posteriors, weighted by observed cost per usable row. Every threshold is relative, so it adapts to a new vertical without retuning.
• Durable execution for long-running agent jobs: leases, checkpointing, pause and resume, idempotent replay after a worker restarts.
• Thirty-plus providers behind one namespaced tool interface, with contact enrichment running as a waterfall and every call cost-tracked against a credit balance.
• Backed by NVIDIA Inception, AWS Activate and Microsoft for Startups.

The company failed on positioning, not engineering. We built horizontally, a dataset for anything, which made it harder to sell and harder to build than the vertical version would have been. The site is still up at datasetlabs.ai, and the worker is at github.com/dylanzonix/datasetlabs-worker.

Those are the first seven weeks. Scenarios were written by users, and the most popular ones drew hundreds of thousands of comments.

• Supervised fine-tuning then DPO on Llama 3.3 70B under DeepSpeed ZeRO-3. Full weight, no adapters, 4096 token sequences, about seven hours a run.
• Regenerate-and-swipe was the core interaction, so every user choice was a labeled preference pair. Four successive preference datasets came out of production traffic.
• A moderation failure tanked the store rating. I trained a 3B scalar scorer to rate generations, then instead of hard-rejecting flagged output, which users experienced as the model breaking, folded the score into the preference data so the policy stopped producing bad output in the first place.
• Fine-tuning made the model repetitive, which measurably hurt retention in a creative product. Temperature did not fix it. Spherically interpolating the fine-tuned weights back toward base did.
• Served it all myself: vLLM for language and speech models, a throughput-optimized custom path for image generation, on AWS and Azure.