B.S. Data Science & AI -- Florida International University
1. Comma.ai's lossy_video_compression_challenge - former #1 (current #3)
Challenge: compress a 1 minute 37.5 MB dashcam video as much as possible while preserving as much detail for SegNet to still identify the road boundaries and PoseNet to determine temporal dynamics - spoiler one of them is lazy and the other is half-blind.
Ranked #1 of 64 on the comma.ai video compression leaderboard, score 0.172 on comma's official T4 evaluation over 600 samples.
| Rank | Entry | Score (lower is better) |
|---|---|---|
| ►1 | semantic-pose-HPAC_CPR1 (mine) |
0.172 |
| 2 | rhnerv_latent_polish |
0.187 |
| 8 | hnerv_ft_microcodec (prize winner) |
0.193 |
| 9 | hnerv_lc_ac (prize winner) |
0.195 |
| 10 | hnerv_lc_v2_scale095_rplus1 (prize winner) |
0.195 |
My submitted PR extends jas0xf's merged PR #86, which established the semantic-token and HPAC codec family in this challenge. I trained three components for it: a width-96 semantic renderer, a learned low-rank pose basis rendered per frame by one einsum rather than by a network, and an integer HPAC entropy model whose arithmetic-coded token stream is 116,980 bytes of the 191,052-byte archive. What is original in that lineage is exactness: HPAC inference on a bounded integer lattice with cross-device symbol verification, a standalone pose carrier in place of PR #86's NeRV slave renderer, and a deterministic repack gated on the archive hash. The reproducible training recipe and archive rebuilds byte-exact from a clean clone, another entrant reproduced it on their own RTX 5070 at 0.171.
Challenge was active at the time of submission but the prizes deadline had already passed so they went to entries that landed earlier, so this holds the top score and no prize.
My prize was a submission that slashed the former leaderboard's #1 by 8% in a single leap (0.187 → 0.172) after months of downstream entries fighting for 0.001 crumbs. Prior entries spent months trading micro-optimizations for a cumulative 4% gain, my submission doubled the progress of the previous 30 submissions combined.
Five fixes landed on main across three external repositories.
| Repo | Stars | Fix | Landed as |
|---|---|---|---|
tokio-rs/topcoat |
4,030 | formatter panic on nested macros | PR #273 |
stephenschoettler/hermes-lcm |
929 | HTTP 400 from a leading assistant turn | PR #280 |
NousResearch/hermes-agent |
223,369 | signed-reasoning replay crash loop | commit 64628ea |
NousResearch/hermes-agent |
223,369 | WhatsApp senders dropped by alias resolver | commit 263ffec |
NousResearch/hermes-agent |
223,369 | HTTP 400 after a second compaction | commit e71c113 |
The hermes-agent commits landed through maintainer cherry-picks from my closed pull requests, with git authorship preserved. I filed the topcoat panic as #272; the fix merged 57 minutes after I opened the pull request.
19,786 all-time downloads of my gpt-oss-20b tool-use finetune on NousResearch's dataset using a standard unsloth notebook, exported to GGUF, MLX, 4-bit, and 16-bit (Hugging Face).