Video synchronization for watch parties: NTP-style clock estimation over the socket acknowledgment, predictive drift control, room state in Redis behind atomic Lua scripts, and a control protocol model-checked in TLA+. Multi-instance-capable (exercised in a three-instance lab and nightly CI; production runs one backend instance). Go and Rust relays exist as conformance studies.
48 ms P95 steady-state drift · 3 Chrome clients · ~300 ms simulated RTT · 240 splayer-reported, single committed run · docs/measurements/servo/S2-overhauled-servo-mse76wbj
TypeScriptNestJSReactSocket.IORedisPostgreSQLTLA+GoRust
Fused residual-add + normalization kernels for PyTorch with FP8 outputs and a deterministic backward, registered as custom ops so torch.compile keeps the graph whole. A100 and H100 results are committed, including the configurations where PyTorch wins.
1.04–1.76× vs the torch.compile composite · RMSNorm→FP8 · A100 · FP16 · kernel time4.9–7.2× vs eager; H100 shows 0.70–1.20× and makes no claim · benchmarks/results/2026-08-25_a100-40gb_v050_ops
CUDAC++PyTorchtorch.compileFP8pytest
A training autotuner that times precision, compile mode, memory format and fused-optimizer configurations on the target model and batch, keeps the fastest, and never returns one slower than the baseline. Version 2.0 on PyPI; the 1.0 marketing numbers were retracted in the changelog.
2.7–6.7× training-step speedup vs PyTorch-default FP32 eager · A100ResNet-50 2.71×, ResNet-18 3.84×, 6-layer Transformer 6.65× (2.55× vs TF32 eager) · benchmarks/results/2026-08-25_a100-40gb_v200
PythonPyTorchCUDA graphsAMPPyPI
A FlashAttention-style Multi-Query Attention CUDA kernel for PyTorch: shared-memory tiling, online softmax, FP32 forward, checked against a float64 reference. It beats my earlier kernel by 17–20× and still loses to PyTorch's tensor-core SDPA.
0.25–0.40× the speed of F.scaled_dot_product_attention FP32 · A100 · max error vs float64 ≤ 8.9e-7benchmarks/results/bench_2026-08-20_210834_a100-sxm4-40gb.json
CUDAC++PyTorchpytest
A Donut vision-encoder-decoder fine-tuned to transcribe synthetic handwriting-style prescriptions, published on Hugging Face as a model and a 1,000-image dataset, with a Gradio demo that gates OCR behind zero-shot prescription classification. Not evaluated on real handwriting.
6,982 all-time model downloads · 652 in the last 30 days · 8,060 all-time dataset downloads · 24 GitHub starshuggingface.co/chinmays18 · fetched 2026-10-07 · no accuracy figure is quoted because none is committed
PyTorchTransformersDonutLightningGradio
A self-hosted API development platform: collections and environments, mock servers, rule-based OpenAPI governance, natural-language test generation and a drag-and-drop workflow builder, with a live demo and a published VS Code extension. The AI-assisted analysis is optional.
307 REST endpoints on a fresh local install · 97 unit tests · 447 VS Code Marketplace installsREADME measured-numbers table, 2026-08-25 (the live demo's openapi.json exposes 265 path/method pairs) · Marketplace gallery API, fetched 2026-10-07
FastAPIReactSQLAlchemyWebSocketsDockerTypeScript
product · chrome web store
A published Chrome extension that adds real-time prompt coaching and searchable conversation memory to ChatGPT, backed by a FastAPI service with Supabase storage and OpenAI embeddings, plus a React dashboard. Built over one week in July 2025 and still live.
4.9 / 5 from 12 ratings · 23 users · listing liveChrome Web Store, fetched 2026-10-07 · small sample
Chrome MV3FastAPISupabaseOpenAI APIReactTypeScript
A C++17 / OpenGL 3.3 viewer with an octree for frustum culling and distance-based level of detail, an XYZ-RGB loader, voxel downsampling and a free camera. Renders generated files of up to 2 million points. The frame-rate figures it used to advertise measured CPU submission time only, so none are repeated here.
Octree frustum culling + LOD · files up to 2 million points · CMake build verified 2026-10-07no frame-rate claim; the old 858 FPS figure came from a 100,000-point run
C++17OpenGLGLSLCMake
A Flask API that builds a conflict-free timetable for 500 synthetic students with Google OR-Tools CP-SAT: sections, capacity, time conflicts, a 3–5 course load and prerequisites as hard constraints, with a preference-weighted objective. The solver runs under a configured 60-second cap, so the schedule is feasible rather than proven optimal.
500 synthetic students · 1,880 assignments · 0 conflicts by constructionsingle committed run, full_metrics.json; not reproducible from the committed seed (capacities 30–60 in seed_data.py vs 150–250 in the results) · zero conflicts is a hard constraint, not a result
PythonOR-ToolsFlaskPostgreSQLDocker
A full-stack prototype that detects walls in uploaded floor plans with OpenCV and runs a genetic algorithm over a physics-based RF path-loss model (free-space loss, ITU-R P.1238 indoor model, material attenuation, knife-edge diffraction) to propose Wi-Fi router placements and render coverage heatmaps.
25 router models in the seeded catalog · 971-line genetic algorithm · 1,232-line RF propagation moduleprototype: no tests, CI or deployment, so no accuracy or coverage figure is quoted; of ~17k Python lines, ~13.5k are alternative wall detectors the default mode never runs
FastAPICeleryOpenCVPostGISReactVite
Not listed: gradient-cache, multi-ai-browser-extension, llm-knowledge-assistant and marketing-attribution-optimizer. They remain public on GitHub, but their headline claims did not survive a check against what is committed (one has never compiled), so they are not presented here as work.