High-performance AI infrastructure, GPU compiler design, temporal memory benchmarks, and distributed systems written in Python, C/Metal, and Go:
- LongArc — Open, reproducible benchmark for temporal memory, privacy boundaries, and safe proactivity in long-horizon AI agents. 1,200 synthetic scored queries across 15 task families. 33/33 unit & property tests green.
- Cairn — Durable, policy-aware delivery runtime with transactional reservations, outbox persistence, worker lease fencing, and simulated CDC reconciliation in Go. Full unit & outbox test suite green.
- TileForge — Native Metal GPU Tensor Compiler. Translates high-level SSA IR into C++ Metal Shading Language (MSL) with threadgroup shared-memory tiling, SIMD barriers, and semantics-preserving optimization passes. 71/71 tests green.
- Oriel — Evaluation-driven model router & benchmark suite. Employs Thompson sampling over Beta posteriors, request hedging, and Wilson score confidence bounds to dynamically route traffic under latency/cost constraints. 54/54 unit tests green.
- Recall — Bitemporal entity and fact memory engine for long-horizon agents. Combines point-in-time state reconstruction, half-open validity intervals, and concurrency stress testing. 32/32 unit tests green.
- PagedServe — High-throughput LLM inference systems reference featuring block-based PagedAttention KV-cache management, continuous batching, prefix caching, and speculative verification. 290/290 unit tests green.
| Repository | Domain | Core Tech | Test Suite Status | CI Workflow |
|---|---|---|---|---|
| LongArc | Temporal Agent Benchmark | Python 3.11+ / Pydantic / DuckDB | ✅ 33/33 Passed | GitHub Actions |
| Cairn | Transactional Outbox & Delivery | Go / PostgreSQL | ✅ Passed | GitHub Actions |
| TileForge | GPU Compiler & IR | C++ Metal MSL / Python SSA | ✅ 71/71 Passed | GitHub Actions |
| Oriel | Model Router & Evals | Thompson Sampling / FastAPI | ✅ 54/54 Passed | GitHub Actions |
| Recall | Bitemporal Memory Engine | FastAPI / SQLite WAL / Pydantic | ✅ 32/32 Passed | GitHub Actions |
| PagedServe | LLM Inference & KV Cache | PagedAttention / PyTorch | ✅ 290/290 Passed | GitHub Actions |
