A data science project that transcribes, analyzes, and ranks every 1-minute stand-up set from the Kill Tony podcast — and uses that data to answer the question: what is the best Kill Tony minute of all time?
Kill Tony is a weekly stand-up comedy podcast hosted by Tony Hinchcliffe. Random comedians pull a golden ticket from a bucket and get exactly one minute on stage. After the set, Tony interviews them live. The show has run for 700+ episodes with some of the biggest names in comedy as guests.
It's one of the most unique formats in stand-up — raw, unfiltered, and ruthlessly honest. A lot of careers have started (and ended) on that stage.
Every episode gets run through a two-pass AI pipeline:
-
Pass 1 — Transcription: Episode audio is downloaded and fed to Gemini in overlapping 20-minute chunks. Each chunk comes back as a speaker-labeled transcript (
tony,comedian:hans_kim,crowd, etc.) -
Pass 2 — Analysis: The merged transcript text is sent back to Gemini for structured data extraction: comedian name, set transcript, topic tags, crowd reaction, Tony's praise level, whether they got a golden ticket, interview summary, and more.
-
Kill Score: Each set gets scored based on crowd reaction, Tony's praise, joke book size, and bonus events (golden ticket, secret show invite, told to sign up again).
-
Frontend: A SvelteKit dashboard at killtonyarchive.com surfaces the data — top sets, topic trends over time, guest appearances, episode pulse charts, and a side-by-side voting system.
| Layer | Tech |
|---|---|
| Pipeline | Python, Google Gemini API (audio + text), yt-dlp, ffmpeg |
| Backend | FastAPI, SQLite |
| Frontend | SvelteKit, TypeScript |
| Hosting | Railway |
YouTube URL
│
▼
yt-dlp → MP3
│
▼
Split into 20min chunks (3min overlap)
│
▼
Pass 1: Gemini Audio → Speaker-labeled transcript JSON
│
▼
Merge chunks, deduplicate overlaps, normalize speaker labels
│
▼
Pass 2: Gemini Text → Structured set data (JSON)
│
▼
SQLite → FastAPI → SvelteKit
The two-pass design was deliberate — single-pass (transcribe + analyze together) produced sparse, low-quality results. Separating concerns gave 10x more transcript entries per episode and dramatically better set extraction accuracy.
kill_score = (tony_praise_level × 2) # 2–10 pts
+ crowd_reaction_score # 0–4 pts (silence → roaring)
+ joke_book_score # 0–3 pts (none → large)
+ golden_ticket × 10 # bonus
+ secret_show_invite × 5 # bonus
+ sign_up_again × 2 # bonus
Theoretical max: 29. Used to rank every set in the database.
backend/
batch_processor.py # Main pipeline (download → Pass 1 → Pass 2 → SQLite)
reprocess_pass2.py # Re-run Pass 2 only (uses cached transcript)
daily_processor.py # Incremental episode processor for scheduled runs
main.py # FastAPI server + REST endpoints
database.py # SQLite schema, CRUD, kill score computation
episodes.json # Episode list with processing status
frontend/
src/routes/ # SvelteKit pages (overview, episodes, sets, guests, vote)
src/lib/api.ts # API client
data/
transcripts/ # Cached Pass 1 transcript JSON (gitignored)
scripts/
sync-db-to-railway.sh # Upload local DB to production
A running log of architectural decisions and why they were made is in DEV_LOG.md.
Some highlights:
- Two-pass pipeline instead of single-pass — single-pass outputs were too sparse to be useful
- Chunked audio (20min + 3min overlap) — Gemini hits output token limits on full-length episodes (~90min), silently dropping the last 30+ minutes
appearance_numnot extracted by AI — Gemini hallucinated these values; now computed from the full DBjoke_densityremoved from kill score — too noisy; Gemini joke counts are subjective and timestamp drift affects set window accuracy
- Transcription timestamps from Gemini can drift 1–2 minutes from actual audio position (a known limitation of LLM audio transcription)
- The planned fix is replacing Pass 1 with local WhisperX for word-level timestamps
- Sentiment/demographic fields were explored and removed — unreliable at scale
See FUTURE_ENHANCEMENTS.md for the full list. Top candidates:
- WhisperX for accurate timestamps → enables laughter attribution per set
- Joke-level segmentation (score individual jokes, not just full sets)
- Comedy DNA clustering (k-means on topic tags to group comedians by style)
- Survival analysis (what % of bucket pulls ever return?)
- YouTube comment sentiment vs AI kill scores
This started as a portfolio project but the data is genuinely interesting. If you're a Kill Tony fan and want to contribute — better prompts, additional episode coverage, new visualizations — PRs are welcome.
See CONTRIBUTING.md for setup instructions.
If you enjoy this project, consider buying me a coffee:
MIT