I build ML systems and the backends around them.

Problem. A NeRF learns a 3D scene from a set of photos. If an attacker can edit some of those photos, can they remove one object from the scene, and what does that do to everything else? To measure it I built a Blender pipeline for a procedural scene and camera rig, a scripted compositor that erases the target from selected views, retraining, and evaluation of the target region and the rest of the frame separately, on a held-out view set that was frozen and checksummed before the first poisoned run.

Key decision. The methodology was locked before any results existed, and every non-trivial choice since is logged with the alternatives I rejected. That discipline paid off before the main sweep: it caught four silent config defects, including every poisoned condition quietly training on clean data, before any compute was spent on them.

Result. The proof of concept passed: target‑region PSNR fell from 23.6 to 19.4 to 11.6 dB as the poisoned share went from 0 to 20 to 50% (Figure 1). Next is an 8-condition, 3-seed sweep at 200,000 iterations per run.

Problem. Give it a GitHub repository or a product description, and it places the AI system in an EU AI Act risk tier, with UNESCO and IEEE ethics flags. Asking an LLM whether a system is high-risk gives an answer nobody can audit, and it can change between runs.

Key decision. The LLM only extracts facts into a fixed schema, quoting the text as evidence for each one. An ordered table of plain Python rules makes the decision: the first match wins, and the same facts always give the same tier, with the rule and legal provision that decided it.

Result. Deployed and public, with 102 tests.

Problem. Fine-tuning a pretrained model means a lot of judgment calls: which model, which layers to freeze, what learning rate. LayerWise is meant to look at an image dataset and make those calls with a reason attached to each one.

Key decision. The engine runs independently of the API, so every decision can be tested without a server.

Result so far. A dataset analyzer that computes per-class statistics over sampled images, and a domain detector that tells natural photos from medical, satellite and other imagery, with a confidence score and a readable chain of evidence. The recommender itself is not built yet; it is next.