Systems I've designed, built, and shipped.

A hardened service that ran 24/7 on a single server — ingesting real-time data from multiple external sources (streaming and REST), fusing LLM inference with a 460k+ record historical corpus to produce retrieval-grounded decisions on a ~110-second loop, then executing actions automatically through signed API calls, with staged decision gates, risk-aware sizing, and full per-cycle observability.

• A self-healing engine time-aligns two independent real-time data sources simultaneously.
• Run as pre-registered, falsifiable experiments — each hypothesis instrumented and measured against real outcomes, then kept or retired on the evidence — producing a reusable measurement instrument and a public 462k-row dataset.
• Disciplined delivery: versioned releases with soak windows, codified smoke tests, and static-analysis gates.

Runs a multi-stage workflow end to end with no human in the loop per cycle — it sources and enriches inputs, evaluates each against an evidence-based scoring engine, generates personalized content with an LLM, sends compliant email and classifies the replies, processes payments through Stripe, and auto-delivers generated documents.

• Adaptive workload allocation via Thompson sampling decides where to focus effort with no human input.
• At-least-once payment semantics: revenue is recorded before the irreversible delivery step, so a crash never loses accounting.
• Production-grade safety and compliance: SSRF defense on a public endpoint, regulatory email compliance, and per-domain deliverability/reputation management.

Monitors multiple external feeds, normalizes them through a source-agnostic intake layer, scores each item for relevance with an LLM, and for qualifying items auto-generates tailored documents — dropping the finished package into a review queue for one-tap human approval. It never acts on the irreversible step on its own.

• Source-agnostic intake: every feed normalizes to one internal record behind interchangeable adapters, so new sources plug in without touching the core.
• LLM scoring with a deterministic spine — a validator clamps outputs and routes low-confidence items to human review.
• Every generated draft passes a voice check and a compliance check before it's queued.