Projects

What happens when a website is under attack, but blocking the attacker might also block real customers?

Portcullis helps investigate the attack and propose a solution. An operator describes what's going wrong, and an AI agent examines the available evidence, identifies a likely cause, and drafts a firewall rule to address it. Before recommending the rule, Portcullis tests it against recorded network traffic to estimate how much malicious traffic it would stop and how many legitimate users it might affect. It also compares its proposal with a simpler, baseline rule so the operator can evaluate the trade-offs.

Built on Cloudflare Workers, this project uses Durable Objects to maintain incident state and Cloudflare Workflows to coordinate the investigation, including retries and pauses while awaiting approval. It uses Llama 3.3 70B through Workers AI to reason about incidents and propose rules, while deterministic code handles validation and traffic analysis.

I evaluated Portcullis using a simulated traffic environment with eight scenarios, including five deliberately designed to expose a weakness in naive firewall rules: blocking the attack can also block legitimate users. The initial real-model evaluation is limited to two runs, so the results are preliminary.

I originally built Oblique as an AI writing system, but eventually ran into a problem I couldn't prompt my way around: regardless of what I asked for, the output kept falling back into the same recognizable "AI voice." So I started over.

The current version uses a five-stage generation pipeline informed by research into what makes generated fiction feel synthetic. I deliberately built directly against the model API rather than using an orchestration framework, giving me visibility into every stage of generation, while SQLite telemetry makes failures and intermediate outputs inspectable. An 11-prompt evaluation harness compares the pipeline against a same-model baseline, so improvements are measured rather than judged entirely by feel.

Instead of relying on a single retrieval method or a managed RAG service, I built a hybrid retriever over 1,059 Wikipedia pages and 10,955 chunks, combining dense vector search, BM25, Personalized PageRank, and entity resolution. The idea was simple: each retrieval signal sees information differently, so their failure modes should be complementary.

I built an evaluation framework using Recall@k and MRR@k to test that assumption, with the resulting system showing a 38% relative diversity gain on multi-hop retrieval. I also profiled the same pipeline across Apple Silicon and CUDA to understand its behavior across hardware backends.

Small sensors struggle in difficult lighting, so Nova Bloom captures multi-exposure RAW and JPEG bursts of up to 100 frames along with exposure and motion metadata for on-device processing. I designed the processing pipeline around ideas from recent computational photography research, including exposure normalization, frame alignment, motion rejection, multi-frame fusion, and RAW finishing.

The full multi-frame pipeline is implemented but currently disabled while I continue tuning its numerical behavior. Rather than ship an unreliable result, the app currently uses a simpler single-frame RAW finishing path. Building it has made me think as much about concurrency, performance, and reliability as photography itself.

VERA combines vision, speech, personal knowledge retrieval, and language-model reasoning through a Swift iOS client and an asynchronous Python backend. Rather than relying on a proprietary multimodal API, I built a modular late-fusion pipeline that processes vision, speech, and retrieval independently before assembling the context for a Llama model.

The MacBook Pro powering the client side was a CPU-bound 2018 Intel model, well short of what real-time vision, speech, and LLM inference actually need, so I split the system accordingly: perception and I/O ran at the edge, while the orchestration layer and model inference ran on Azure. I built a 50-query multimodal benchmark to evaluate the system, measuring latency, retrieval quality, and task completion, and used ablation tests to understand how much vision, speech, and retrieval each contributed to the final result.