I'm a computer science undergrad at UGA. Most of what I build is the layer underneath machine learning — the schedulers, indexes, collectives and evaluation harnesses that decide whether a model is actually usable — written from scratch and benchmarked against the thing it is supposed to beat.
Two rules I hold myself to across every repo here:
- The deliverable is a measurement, not a demo. Every project ships a benchmark you can run, against a real baseline, on data I didn't generate to flatter myself.
- Report where it loses. A speedup with no regime where it's slower is usually a broken benchmark. Each README has a section on the cases the approach handles badly, because that's the part that shows I understand it.
| project | what it is | the number that matters |
|---|---|---|
| raft-mvcc | C++17 distributed-database correctness lab — Raft elections/log repair/majority commit, serializable MVCC, and P-compositional linearizability checking | 598 assertions across seeded partitions; ~0.69M 3-node replicated writes/s and 11-tick five-node failover |
| columnar-engine | C++17 vector-at-a-time engine — null bitmaps, selection vectors, joins, aggregation, TPC-H Q1 and scalar differential oracles | 6,288 assertions; Q1 processes 53.1M rows/s and batched filters reach 1.94× scalar |
| lsm-tree | C++17 storage engine — WAL, memtables, SSTables, Bloom filters, version-aware range scans and crash-safe compaction | 4,064 assertions plus 100 randomized hard crashes with zero acknowledged writes lost |
| wal-recovery | Transactional WAL — CRC32 records, atomic replay, damaged-tail repair, external crash oracle and validated group commit | Zero crash-campaign losses; 16-way group commit reaches 40.2k tx/s, 2.58× single commit |
| vllm-lite | LLM inference server — paged KV cache, continuous batching, prefix caching, speculative decoding, OpenAI-compatible API | 94% KV utilization vs 21% for static batching; 2.9x throughput at 5.5x better TTFT |
| annlite | HNSW vector index in C++17 — hand-vectorized distance kernels, Python bindings, HTTP service | Parity with FAISS at 0.99 recall, 1.83x faster at 0.999 — the same Pareto frontier, not a beaten benchmark |
| dist-train | Distributed training — ring all-reduce built from send/recv, plus data, tensor and pipeline parallelism |
Ring moves 58.7 MB/worker vs 234.9 MB naive at 8 workers, and the gap widens exactly as the arithmetic predicts |
| rag-eval | RAG pipeline plus a reproducible eval harness that gates retrieval quality in CI | BM25 on SciFact reproduces the published BEIR number — 0.6643 vs 0.665 — so the ablations measure ideas, not my bugs |
| feature-store | Point-in-time correct training data, streaming materialization, drift monitoring | Time-travel leakage inflates the offline score by 0.059 AUC, measured by keeping the leaky join next to the correct one |
| grammar-decode | Constrained decoding — JSON Schema compiled to a character automaton, then a cached token mask over a 50k vocabulary | 100% schema conformance against a 0% baseline; masking up to 1203x faster than replaying the vocabulary |
| agent-harness | LLM agent runtime — a five-layer sandbox and a programmatically graded benchmark | 57 tests, 25 of them escape attempts. The sandbox is tested by attacking it, and refuses to run where it can't enforce |
| codebase-qa | Codebase Q&A service built around the unglamorous parts: auth, rate limiting, durable job queue, cost tracking, tracing | scrypt-hashed keys, constant-time comparison, revocation as a timestamp — the prototype-to-service gap, written down |
| sdr-receiver | QPSK software-defined-radio receiver — RRC pulse shaping, carrier/timing acquisition, soft decisions, regular LDPC decoding | Full impaired receiver tracks coherent theory; 0 observed errors in 21,600 bits at 4–8 dB after the LDPC waterfall |
| track-fusion | Multi-target tracking — IMM filter bank, JPDA data association, Wald-test track scoring, OSPA evaluation | IMM cuts localisation error 47% through a manoeuvre and costs 45% on a target that never turns — both ends measured |
| sar-focus | SAR image formation — pulse compression, backprojection, phase gradient autofocus, impulse response analysis | Resolution within 0.5% of 0.886·c/2B and 0.886·λR/2L; sidelobes converge to the −13.26 dB theoretical floor |
| vio-nav | GPS-denied navigation — MSCKF visual-inertial odometry, IMU preintegration on SO(3), null-space feature marginalisation | 51.9x lower drift than inertial dead reckoning at 40 s; 65x when the IMU bias starts unknown |
Regenerated daily by a GitHub Action that reads every repo and counts the test functions in the source — not typed in by hand, and not fetched from a third-party stats service that could rate-limit or vanish. It reports test functions; parametrized suites and looped invariants expand those into many more checks (598 assertions in raft-mvcc alone).
Not sixteen unrelated weekend projects. Twelve form one path through a production ML stack and the storage substrate beneath it. The other four are sensing and estimation — coherent imaging, target tracking, recovering bits from a corrupted waveform, and navigating without GPS — which is the same habit of measuring against a known answer, pointed somewhere harder to fool yourself about.
flowchart LR
subgraph STORE [store]
M[wal-recovery<br/>atomic durable replay]
N[lsm-tree<br/>SSTables + compaction]
P[raft-mvcc<br/>consensus + snapshots]
O[columnar-engine<br/>vectorized analytics]
end
subgraph TRAIN [train]
B[feature-store<br/>point-in-time joins]
A[dist-train<br/>ring all-reduce]
end
subgraph SERVE [serve]
C[vllm-lite<br/>paged KV + batching]
D[grammar-decode<br/>constrained decoding]
end
subgraph RETRIEVE [retrieve]
E[annlite<br/>HNSW index]
F[rag-eval<br/>eval harness]
end
subgraph APPLY [apply]
G[agent-harness<br/>sandbox + benchmark]
H[codebase-qa<br/>the boring parts]
end
subgraph SIGNAL [sense and receive]
I[sdr-receiver<br/>QPSK + LDPC]
J[sar-focus<br/>SAR image formation]
K[track-fusion<br/>IMM + JPDA tracking]
L[vio-nav<br/>MSCKF visual-inertial]
end
M --> N --> P --> O --> B --> A --> C --> D
E --> F --> H
C --> G
D --> H
I --> J --> K --> L
systems C++17 · Python · Java · Raft · MVCC · WAL · LSM/SSTables · columnar execution · SIMD · POSIX
ml PyTorch · torch.distributed · HNSW · BM25 · cross-encoder reranking
signal QPSK · RRC filters · carrier/timing acquisition · LDPC · min-sum decoding
radar SAR backprojection · PGA autofocus · Kalman/IMM filtering · JPDA · OSPA
estimation IMU preintegration · SO(3)/Lie groups · MSCKF · ATE/RPE · triangulation
services FastAPI · SQLite/Postgres · durable job queues · scrypt · tracing
industrial Power BI · Node-RED · CtrlX Core · ERP automation
Day job: automation and analytics at MP Equipment — dashboards, ERP automation, and AR/AI for industrial food processing.
aaryansp26@gmail.com
Every repo above is MIT licensed. Clone one and run the benchmark — that's what it's there for.
Banner generated by banner.py: animated SVG, two themes, respects prefers-reduced-motion.