Haoran (Casper) Zhang

A deterministic, invariant-checked discrete-event simulator for accelerator fleet scheduling — topology-aware gang placement on a TPU-style 3D torus, priority preemption, quotas, scripted failures — paired with a real multi-process controller/agent runtime over gRPC that detects failures by heartbeat timeout, reassigns work under epoch fencing, and proves its recovery claims in asserted demos. A FastAPI experiment layer and a Next.js dashboard drive it; benchmarks are reproducible.

• Wrote a deterministic C++20 discrete-event scheduler that places 100,000 tasks across a 10,000-node TPU-style 3D torus in ~80 s single-process — verified by 381 tests and a 10,000-seed randomized invariant sweep whose vacuousness guard rejects runs that schedule nothing.
• Built a control plane of 20 OS processes (gRPC, Docker) with heartbeat failure detection, epoch-fenced reassignment, and dedup under injected duplicates; a scripted split-brain run asserts the full 8-event recovery chain in order, not just its presence.
• Reported a negative result: priority-gang (topology-aware + preemption) does not complete under contention — its planner re-runs the placement search per candidate victim, so cost is a product of the factors, not their sum. Release-suite figures regenerate byte-for-byte from a retained manifest; every run traces to a scenario hash and seed.

An AI-guided mediation platform that interviews each side privately, transforms unstructured narratives into structured case intelligence, and produces a transparent resolution brief — including uncertainty-aware responsibility analysis. All data stays in the browser; no accounts, no server database.

• Built a 4-stage LLM pipeline — adaptive interviewing, structured extraction, statement generation, and comparative ruling — with outputs validated in code, not prompts: 35 runtime rules across 5 per-stage validators reject bad output before acceptance.
• Enforced no-fallback semantics for AI calls: a failed call stays pending for explicit retry and is never silently replaced by local imitation.
• Kept 100% of case data client-side across 13 typed IndexedDB tables — no accounts, no server database — verified by 784 test cases that run fully network-disabled.

A retrieval-augmented nutrition assistant over 5,907 PubMed abstracts and a pinned USDA FoodData Central snapshot, running entirely on one 6 GB laptop GPU. Two answer paths: literature questions go through retrieval and generation, while nutrient lookups resolve to a SQL query against the snapshot. Built around a frozen baseline and a nine-arm ablation.

• The model never produces a nutrient value — it extracts two strings and SQL does the rest, which is why the deterministic path scores 88.3% (106/120) on a held-out split with 0 out-of-tolerance values and 0 unit errors. Every failure is food-name resolution.
• Froze the baseline's inputs, not just its function: a digest covering the prompt, model tag, top_k, and the PubMed query that decides what is in the index at all. A test fails on any drift, checked before and after each of nine arms.
• Reported a negative result: the improved pipeline (rewrite → hybrid → rerank) measured worse than the frozen baseline it was built to beat — 0.5823 vs 0.7983 across 39 human-reviewed questions. A judge-free measurement found the cause: arms containing the reranker retrieve the known-relevant abstract in 29–30 of 36 questions, arms without it in 34. The demo ships the baseline's retrieval.