Weekend Processing Marathon
I'm deep in the weeds this weekend, and I love it. While the landing page at https://askthegame.felo5.com/ becomes more visually appealing, the real work is happening behind the scenes. I've processed 400+ episodes just today, and I'm racing to get through as many as possible before Monday.
I don't want to mess with the system while it's working. It's like performing surgery on a patient who is still moving. One wrong move and you break something perfectly fine.
So I used this time to craft a better story of what I'm doing on the landing page until I finish this processing marathon. The pipeline is humming along beautifully, confidence reports are generating clean data, and episodes are getting chunked into thousands of semantic segments. Why would I risk that for a feature tweak?
What's Actually Happening
Each episode goes through this gauntlet:
- Voice recognition with speaker diarization
- Semantic topic labelling using AI that understands business context
- Confidence-based filtering that removes unreliable segments
- Quality reporting that tells me precisely what worked and what didn't
The numbers are solid. I'm seeing 91.6% pass rates on confidence filtering, which means the system is becoming increasingly accurate at distinguishing between correct and incorrect predictions.
From today's batch alone:
- 112 new confidence reports generated
- Thousands of segments labelled with business topics
- Episodes ranging from 100+ to 164+ in the sequence
- Consistent quality metrics across different episode types
The pipeline configuration is dialled in. Each run produces detailed reports showing segments passed, flagged, or dropped based on confidence scores. It's not just processing audio - it's learning what good data looks like.
Why This Matters
I started this project with one question: "Can I find when Alex talks about pricing and actually listen to that part?" Now I'm sitting on a system that can identify speakers across episodes, segment content by semantic meaning, and automatically maintain quality standards.
But none of that matters if I only have 50 episodes processed. The real test is scale. Do you think this can work on 900+ episodes? Can it maintain quality when processing varies wildly between early episodes and recent ones? Can the confidence filtering adapt to different audio quality and speaking patterns?
The answer so far is yes, but I need to prove it with volume.
This week, we have no new features. No architecture changes. No database migrations. No clever optimizations that might break something subtle.
The landing page gets updates because it's purely an exercice of explaining what I'm doing and it shows what's already working. But touching the core pipeline? That's off limits until this processing run is complete.
Finish Line Tomorrow?
I want to finish the weekend with as many episodes as possible processed. Maybe all of them. Then I can go back to building new features on top of a foundation I know is solid. However, right now, the most valuable thing I can do is demonstrate that the pipeline can handle real-world variety and volume.
The marathon continues.
– Benoit Meunier