.png&w=256&q=75)
🤓 Latest Submissions
TextSink — Four Voices. One Truth.
TextSink (TS) is a containerized Track 2 captioning agent built end-to-end on Gemma 4 (26B MoE) via a dedicated Fireworks deployment. one model does the visual grounding AND all four caption styles. THE SCORED PIPELINE. ffmpeg samples frames across the whole clip; Gemma 4 extracts scene facts as strict JSON. If grounding comes back empty, TS retries and refuses to caption blind; captions may only use extracted facts, so accuracy holds on any clip. Four tone contracts with distinct comedic mechanisms (broadcast precision, deadpan irony, dev-culture metaphor, warm relatability) yield four captions an LLM judge can blind-sort. The container implements the standard /input/tasks.json -> /output/results.json harness contract, runs clips concurrently (3 official clips in ~30s warm), and if the Gemma deployment is unreachable it cuts over per-task to serverless models — no task ever scores 0. BEYOND THE CONTRACT. The same grounding powers TS CC; real timed closed captions (.srt/.ass + burned video), each style one consistent voice with running gags. AND The Hecklers: two models (Gemma 4 vs gpt-oss-120b) argue about the video turn by turn, every line generated live. STAN & GUS (sarcastic old men), LINT vs VIBE (coder vs vibe-coder), DORIS & PEARL (gossiping neighbors), matching the 3 styles. GEMMA TEACHES GEMMA; AND THE STUDENT WINS. Gemma 4 drafted candidates at high temp, judged its own drafts on accuracy + tone, and the winners became a self-distilled SFT set. Fine-tuned ts-g3-captioner (Gemma-3-27B LoRA) beats its teacher's prompted best-of-3 18-4 head-to-head under a neutral judge, higher on every style (ab_results.json). EVIDENCE, NOT CLAIMS. GEMMA_PROVENANCE.md (where Gemma authored the entry), eval/ABLATIONS.md (designed mechanisms beat generic prompts 11-5 under 2 judges, on the fallback model), submission/fallback_verification (forced foreign-key run — zero empty captions). All demos generated on the official clips. MIT. Python 3.11 + ffmpeg. No framework lock-in.
13 Jul 2026
Legislative Bill Analyzer and Podcast Generator
Legislation is one of the least accessible documents in modern society. U.S. bills span thousands of pages of dense legalese with cross-references that fan out across the U.S. Code. Manually reviewing a 3,000-page bill takes teams of analysts weeks. Important policy decisions disappear inside documents almost no one has time to read. Nota.Lawyer Bill Analyzer turns any U.S. bill PDF into a structured analysis and an AI-generated podcast video, with the political tone and depth dialled to listener preference. Pipeline 1 chunks the bill on TITLE / Subtitle / DIVISION boundaries (200K cl100k tokens per chunk), runs six specialist agents over each chunk (summarizer, USC cross-referencer, pork finder, conflict spotter, podcast-headline generator, headline ranker), and aggregates a canonical report. Pipeline 2 takes the winning headline, generates 19 dialog lines, 19 slide prompts, and 19 motion prompts via Qwen3-30B, renders 19 still slides via Qwen-Image-2512 (gated by a dual-call OCR + judgment critic, up to 15 retries), animates them with Wan 2.2 i2v + LightX2V, narrates via Qwen3-TTS-12Hz-1.7B custom Alex/Jordan voices, lipsyncs talking-head pairs via InfiniteTalk on the same Wan base, and composes three masters - slides-only, all-talking-head, and an alternating hybrid (default). All four Qwen models plus Wan 2.2 stay resident in VRAM the whole run (~150 GB of 192 GB). vLLM Automatic Prefix Caching on ROCm 7.2.3 means the 10 specialist agents share a single prefill of the bill - measured 94.15x APC TTFT speedup on a 99,727-token prefix, 14.8x wall-clock speedup on the 232K-token summarizer, 68.5% sustained APC hit rate. The Build Back Better Act (HR 5376), 2,468 pages and 927,292 cl100k tokens, took 28 minutes cold to a 3-minute hybrid podcast on one MI300X. Open-source, MI300X-native, reproducible end-to-end. Live demo and source linked below.
10 May 2026
