# Adam DePrince - Open Source Projects Website: https://adamdeprince.com/ GitHub: https://github.com/adamdeprince Short LLM summary: /llms.txt Last updated: 2026-07-22 ## Overview Adam DePrince is a Python and C++ systems programmer with a focus on finance, high-performance computing, and systems development. The projects below are open-source systems software projects and experiments, with an emphasis on practical speedups, SIMD, memory efficiency, market-data workflows, and low-level infrastructure. Selected benchmark examples from the homepage: - Goblin Core uses 28.65 RSS bytes per set member at 1M, 36.2% less than the leanest incumbent in the published comparison. - Goblin Core completes a shared-memory request/reply in about 220 ns. - Goblin Store uses 7.4x less RAM at memcached large-object latency. - Goblin Store delivers 68% more throughput than memcached extstore on the published disk-sized benchmark. ## Project Details ### Goblin Core URL: https://goblin-core.dev Goblin Core is a C++23 answer to two Redis-shaped problems: hot state that consumes too much RAM, and latency-sensitive state that spends too long crossing the kernel. It keeps familiar operations for strings, sets, sorted sets, hashes, segmented and Adaptive-PMA lists, sparse arrays, TTLs, counters, Pub/Sub, authentication, and scripting, then specializes the storage and transport beneath them. RESP and opt-in typed SBE work across TLS-capable TCP, native XLIO Ultra TCP, Unix-domain sockets, shared-memory rings, and polled RDMA. This lets Goblin Core act as a compact conventional server, a triple-digit-nanosecond local service, or a low-microsecond network service without changing the Redis-shaped data model. Version 0.10 adds broker-acknowledged Kafka writes with bounded pending-data backpressure, firehose replication of canonical writes, and a native whole-database stream for replica bootstrap. Snapshots, Kafka replay, and the live stream form the recovery path while Kafka remains responsible for durable history. The current same-host suite gives each engine one serving core. A one-million-member set uses 28.65 RSS bytes/member, 36.2% less than the leanest incumbent, while reaching 2.44M RESP/TCP membership checks per second or 6.73M over the SBE ring. The efficient hash uses 48.67 bytes/field, 20.1% less than Redis 8.8. A one-million-member sorted set uses 52.32 bytes/member and reaches 1.06M RESP/TCP or 3.39M SBE-ring ZADDs/sec. The lowest-latency path completes a synchronous HSET or ZADD request and reply in about 220 ns. Across six operations, the shared-memory ring median is 33-40x faster than the fastest RESP Unix-domain-socket result, with 0.24-0.28 us p99. Native XLIO Ultra runs seven Redis commands at 4.98-8.37 us median over 100 GbE. The polled one-sided RDMA ring was also measured across two Dell R820 servers on a generation-old 40 Gb/s FDR10 InfiniBand fabric. A 500,000-sample, depth-one run measured complete Goblin PING request/reply p50 latency of 2.495 us for SBE and 2.673 us for RESP2. That is only about 0.6 us beyond the same server's two-hop QPI path, putting cross-host Goblin operations close to the latency the four-socket machine already pays to cross its own host fabric. Goblin Core also replayed 14.3 billion real Lichess rating updates into one leaderboard containing 21.6 million players, matched Redis member-for-member and score-for-score. At 763.8 MB resident memory, it used less RAM than every compared server: Redis 7.2.4, Redis 8.8, Valkey 9.1, and Dragonfly. Large blobs remain the domain of Goblin Store. ### Goblin Store URL: https://goblin-store.dev Goblin Store is a C++23 NUMA-aware large-object cache, HTTP server, and streaming caching proxy for values from 256 KiB into the multi-MiB range. It speaks memcache over TCP or native InfiniBand/RoCE and serves HTTP/HTTPS with ranges, ETags, and origin-mirror operation. It keeps the latency-sensitive head in RAM, streams the warm middle from SSD, and reads the cold tail ahead from optional HDD. HugeTLB-backed heads, fixed pools, backpressure, O_DIRECT, Linux io_uring, and core-local data paths keep memory bounded as objects grow. On the published network benchmark, Goblin Store matched memcached's large-object latency while using 7.4x less RAM. With a disk-sized working set on the same HDD, its three-tier configuration delivered 118 QPS versus 70 for memcached extstore, a 68% throughput lead, while retaining the full working set while extstore shed 41.5% under write load. The illustrative storage model puts it at about 25% lower monthly cost. A separate native InfiniBand report reaches 38.2 Gbit/s with 7.7 us median header time-to-first-byte. Small values are explicitly memcached's turf; Goblin Store is a large-object engine. ### stride-align URL: https://github.com/adamdeprince/stride-align stride-align implements Smith-Waterman and Needleman-Wunsch alignment. It is reported as 1.75x faster than Parasail on AVX-512, 1.37x faster on AVX2, 7.5x faster by geometric mean on LoongArch LASX, and 22.4x faster for specific workloads. It can compare strings, lists of numbers, or arbitrary Python objects. ### massive-speedup URL: https://github.com/adamdeprince/massive-speedup massive-speedup is a high-performance Massive.com flatfile parser, data store, and market microstructure simulator. Flatfile iteration is 5x faster than using Python's gzip and csv modules. Its tick-forward iterative design has an integrated latency-aware market simulator. Coupled with rtta's tick-driven indicators, it is intended to make backtesting resistant to common sources of "crystal balling" that can appear in traditional Pandas and bulk-data strategy development. ### rtta URL: https://rtta.dev rtta is a low-latency incremental technical-analysis toolkit. It includes traditional indicators, Kalman-filter-based indicators, and experimental indicators based on recent research. The library supports both batch processing and efficient real-time tick-based interfaces. It is designed for users who do not want to rewrite Pandas-style models to work with live tick data. Coupled with massive-speedup, it encourages tick-by-tick strategy design that can be carried forward toward live deployment. ### negcycle URL: https://github.com/adamdeprince/negcycle negcycle finds profitable FX cycles, identifying the most profitable cycle in 1.8 us on an 11th-generation Intel i7. It is a SIMD-optimized Python library for finding profitable execution chains in currency markets. Internally, it implements dense Bellman-Ford search with SIMD optimization for Intel, Arm, and Loongson. ### fast-kalman URL: https://github.com/adamdeprince/fast-kalman fast-kalman is a C++ and Python Kalman filter implementation. It is reported as 125x faster than OpenCV's Kalman filter on a typical single-price tracker, with peaks of 235x on harder filters. The library supports a wide variety of Kalman filter types, non-uniform update times, and custom certainty values per update. It can also generate tuning parameters from a chunk of input data. Across the full range of filter types, the C++ Kalman filters run 7x-235x faster than OpenCV's. At the Python level, the reported speedup is about 5x, with Python binding overhead accounting for much of the remaining cost. ### DragonArray URL: https://dragon-array.org DragonArray is a Loongson optimization lab for installable LoongArch64 wheels, source patches, benchmarks, and low-level utilities. Its first release completed missing pieces needed to enable NumPy's Loongson 256-bit LASX SIMD path. It achieved a reported geometric-mean speedup of 3.3x over the LSX-targeted code in NumPy 2.5.0.dev0 on a 3A6000 processor. The work builds on NumPy's existing LoongArch SIMD infrastructure. Adam's part included LASX detection, wiring the remaining LASX primitives needed to enable the 256-bit path, and creating vectorized implementations of transcendental functions. A float16 tanh implementation is reported at 26x the throughput of the scalar implementation it replaced. DragonArray now also ships a BLAKE3 LoongArch SIMD port from the public loongarch-simd source branch. Installable TensorFlow wheels for LASX-capable Loongson 6000-series systems are the next project track. ### bsort URL: https://github.com/adamdeprince/bsort bsort is a fast fixed-record sort library for large data files. It is available as both a C++ executable and a Python library. It has locality-of-reference properties intended to play well with LRU memory eviction and can sort files that are much larger than RAM. The project began as an attempt to beat sortbenchmark.org results from a company laptop. A related drunk-coded version was published at https://github.com/pelotoncycle/bsort and later became part of a startup story. ### mosh patches URL: https://github.com/adamdeprince/mosh This mosh work adds missing features that SSH users commonly expect: SSH-like stream forwarding for local, remote, dynamic SOCKS, and agent use cases through flags such as -L, -R, -D, and -A. The patch series carries protocol, frontend, and state-sync changes needed to move those streams through mosh's roaming terminal session. It also includes a forward-error-corrected bulk file copy, giving mosh users scp-like capabilities on poor links, plus a trained-compressor mode for constrained links such as satellite phones. Testing against an Iridium Go at 2.8 kbps was planned for late May 2026. ### CPython SIMD experiment URL: https://github.com/adamdeprince/cpython/tree/3.14-simd-experimental This CPython 3.14 technology demo adds a runtime SIMD dispatch layer and applies it across everyday interpreter hot paths. Some string operations improve by double-digit factors. The branch accelerates base64 and hex codecs, UTF-8 and ASCII scanning, bytes translation and case folding, JSON loads and dumps string handling, csv.reader field scanning, urllib parsing helpers, and a new C accelerator for html.escape. It also includes an experimental Swiss-table-backed set implementation as a benchmark vehicle for evaluating SIMD metadata scans against CPython's existing set design. The main lesson is pragmatic: the largest wins come from bulk text and codec work where per-call overhead is amortized, while short-string cases are mostly limited by object setup and call overhead. The branch includes benchmark and optimization notes so results are reproducible. ### NumPy LASX experiment URL: https://github.com/adamdeprince/numpy/tree/loongson-experimental Loongson processors have two SIMD modes: 128-bit LSX for embedded devices and 256-bit LASX for larger systems. NumPy supports LSX but not LASX, leaving performance on the table. This proof-of-concept port to LASX reports about 80% gains over LSX and larger gains in places that previously had no vectorization. The experiment is intended to become real NumPy patches. ### goblin-cannon URL: https://github.com/adamdeprince/goblin-cannon goblin-cannon is a low-latency SDR radio implementation aimed at HFT-adjacent cross-Atlantic market-data transport. The site frames the problem as follows: New York is 29 ms from London on EXA Express POP-to-POP links, while the speed of light on a great-circle route is about 19 ms. A radio path needs to receive exchange data, sort competing messages, encrypt, apply CRC and error coding, build a waveform, send data, unpack it, run error correction, check CRCs, and decrypt while leaving enough time for a trader to act. The project is described as a 2.5 ms end-to-end SDR radio implementation built with AES, Viterbi decoding, CRC frame checking, QCI encoding, and oversample-based decoding with AVX2 and AVX-512 support. The project also includes business-side mechanics: market and custom messages, transmission-slot bidding, shadow bids for market data based on volatility, budgets, accounting, and lost-message credits. Cross-Atlantic testing was anticipated for fall 2026, gated on an experimental license. ## Toy Websites ### NAICS Code Finder URL: https://naics.goblinreactor.com/ The NAICS Code Finder takes a plain-language description of a business activity and uses semantic search to rank the closest six-digit industries in the 2022 North American Industry Classification System. It also supports batch matching for text files containing up to 1,000 queries. ### CPV Code Finder URL: https://cpv.goblinreactor.com/ The CPV Code Finder takes a plain-language purchasing requirement and uses semantic search to rank the closest entries in the 2008 Common Procurement Vocabulary. It also supports batch matching for text files containing up to 1,000 queries. ### Sloppy Resume URL: https://sloppyresume.com/ Sloppy Resume is a joke site with real math behind it. It extracts text from two PDF resumes, uses Needleman-Wunsch alignment through stride-align, and scores how suspiciously alike they are with a ten-egg meter. It is a toy, not a hiring system or a tool for consequential decisions. ### DragonArray URL: https://dragon-array.org/ DragonArray is a Loongson optimization lab for LoongArch64 wheels, source patches, benchmarks, and low-level utilities tuned for LSX- and LASX-capable systems. Its first release is a NumPy LASX patch and wheel set for Python 3.12-3.14. It now also ships a BLAKE3 LoongArch SIMD port from its public loongarch-simd source branch, with TensorFlow wheels in progress. ### R820 Field Notes URL: https://r820.goblinreactor.com/ R820 Field Notes publishes current TensorFlow builds for compatible pre-AVX2 x86-64 machines. It documents the CPU flags users need to check and separates AVX1 targets with F16C from those without it, making retired enterprise servers useful for modern machine-learning workloads without pretending they support instructions they do not have. ### Bluesky Pulse URL: https://bluesky.goblinreactor.com/ Bluesky Pulse is a live mathematical sample of Bluesky's public feed. It shows posting volume, a word leaderboard with rates and trend forecasts, language activity, and recent posts crossing the wire. The dashboard is built with C++23 and Wt around one Redis ingest and a shared forecast stream. ## Contact Email: f"{given_name.lower()}.{family_name.lower()}@gmail.com" Phone: f'{2**8}-{2*3**2*37}-{13*101}' GitHub: https://github.com/adamdeprince The email and phone are intentionally encoded on the HTML homepage. The phone expression evaluates to Adam's real phone number.