Slide 1
Slide 1 text
Everything Is Showbiz:
Lessons from a
PHP + AI Side Project
Paul Conroy / @conroyp
Slide 2
Slide 2 text
Paul Conroy
From Dublin, Ireland
Started playing with the web 30+ years ago
(Notepad, Frontpage & Geocities!)
CTO at Square1
conroyp.com / @conroyp
👴
🌍
🇮🇪
Slide 3
Slide 3 text
No content
Slide 4
Slide 4 text
● Interview comedians and
celebrities
Slide 5
Slide 5 text
● Interview comedians and
celebrities
● Ask them what they did
yesterday
Slide 6
Slide 6 text
● Interview comedians and
celebrities
● Ask them what they did
yesterday
● That’s it!
Slide 7
Slide 7 text
No content
Slide 8
Slide 8 text
No content
Slide 9
Slide 9 text
No content
Slide 10
Slide 10 text
What do I want?
Slide 11
Slide 11 text
● Some kind of episode transcription
What do I want?
Slide 12
Slide 12 text
● Some kind of episode transcription
● Searchable
What do I want?
Slide 13
Slide 13 text
● Some kind of episode transcription
● Searchable
● Easily share specific parts
What do I want?
Slide 14
Slide 14 text
● Some kind of episode transcription
● Searchable
● Easily share specific parts
● That’s it!
What do I want?
Slide 15
Slide 15 text
● Some kind of episode transcription
● Searchable
● Easily share specific parts
● That’s it!
What do I want?
Slide 16
Slide 16 text
● Some kind of episode transcription
● Searchable
● Easily share specific parts
● That’s it!
What do I want?
Slide 17
Slide 17 text
● Some kind of episode transcription
● Searchable
● Easily share specific parts
● That’s it!
What do I want?
Slide 18
Slide 18 text
INGEST TRANSCRIBE SERVE
A Simple Plan!
Slide 19
Slide 19 text
Things I already
know how
to do
Challenges in
this project
Audio Transcription
Semantic Search
Dynamic front end
Model evaluation
RSS Parsing
Queues
Cost-effective
hosting
Saying “we’ll tidy that
up later
Convincing myself it’s
“MVP”
Making confusing
Venn diagrams
Slide 20
Slide 20 text
I needed: I get:
Something to run an ingest twice a week Scheduler
Process files in the background Queues
Try different AI tools Service classes & container
A simple UI to browse results Blade, Alpine
Deploy & run at low cost Cache, config, cheap VPS
Slide 21
Slide 21 text
TRANSCRIBE SERVE
INGEST
A Simple Plan
Slide 22
Slide 22 text
Let’s get started!
Slide 23
Slide 23 text
● Design the entities / models
Let’s get started!
Slide 24
Slide 24 text
● Design the entities / models
● Map out the Controllers, Jobs, Services
Let’s get started!
Slide 25
Slide 25 text
● Design the entities / models
● Map out the Controllers, Jobs, Services
● Set up config and environments
Let’s get started!
Slide 26
Slide 26 text
● Design the entities / models
● Map out the Controllers, Jobs, Services
● Set up config and environments
● Set up test / validation rig
Let’s get started!
Slide 27
Slide 27 text
● Design the entities / models
● Map out the Controllers, Jobs, Services
● Set up config and environments
● Set up test / validation rig
● Fix problems with test / validation rig
Let’s get started!
Slide 28
Slide 28 text
● Design the entities / models
● Map out the Controllers, Jobs, Services
● Set up config and environments
● Set up test / validation rig
● Fix problems with test / validation rig
● Build queue infrastructure
Let’s get started!
Slide 29
Slide 29 text
● Design the entities / models
● Map out the Controllers, Jobs, Services
● Set up config and environments
● Set up test / validation rig
● Fix problems with test / validation rig
● Build queue infrastructure
● Fix problems with queues not running
Let’s get started!
Slide 30
Slide 30 text
● Design the entities / models
● Map out the Controllers, Jobs, Services
● Set up config and environments
● Set up test / validation rig
● Fix problems with test / validation rig
● Build queue infrastructure
● Fix problems with queues not running
● ….
Let’s get started!
Slide 31
Slide 31 text
No content
Slide 32
Slide 32 text
No content
Slide 33
Slide 33 text
No content
Slide 34
Slide 34 text
No content
Slide 35
Slide 35 text
No content
Slide 36
Slide 36 text
No content
Slide 37
Slide 37 text
● Design the entities / models
● Map out the Controllers, Jobs, Services
● Set up config and environments
● Set up test / validation rig
● Fix problems with test / validation rig
● Build queue feature
● Fix problems with queues not running
● ….
Let’s get started!
Slide 38
Slide 38 text
No content
Slide 39
Slide 39 text
No content
Slide 40
Slide 40 text
No content
Slide 41
Slide 41 text
TRANSCRIBE SERVE
INGEST
A Simple Plan
Slide 42
Slide 42 text
Whisper
curl --request POST \
--url https://api.openai.com/v1/audio/transcriptions \
--header "Authorization: Bearer $OPENAI_API_KEY" \
--header 'Content-Type: multipart/form-data' \
--form file=@/path/to/file/audio.mp3 \
--form model=gpt-4o-transcribe
{
"text": “Imagine the wildest idea that you've ever had, and you're curious about how it
might scale to something that's a 100, a 1,000 times bigger.....”
}
● OpenAI API
● Fast!
● Supports prompting
● Post-processing
https://developers.openai.com/api/docs/guides/speech-to-text/
Slide 43
Slide 43 text
Whisper
curl --request POST \
--url https://api.openai.com/v1/audio/transcriptions \
--header "Authorization: Bearer $OPENAI_API_KEY" \
--header 'Content-Type: multipart/form-data' \
--form file=@/path/to/file/audio.mp3 \
--form model=gpt-4o-transcribe
{
"text": “Imagine the wildest idea that you've ever had, and you're curious about how it
might scale to something that's a 100, a 1,000 times bigger.....”
}
● OpenAI API
● Fast!
● Supports prompting
● Post-processing
https://developers.openai.com/api/docs/guides/speech-to-text/
Slide 44
Slide 44 text
Whisper
curl --request POST \
--url https://api.openai.com/v1/audio/transcriptions \
--header "Authorization: Bearer $OPENAI_API_KEY" \
--header 'Content-Type: multipart/form-data' \
--form file=@/path/to/file/audio.mp3 \
--form model=gpt-4o-transcribe
{
"text": “Imagine the wildest idea that you've ever had, and you're curious about how it
might scale to something that's a 100, a 1,000 times bigger.....”
}
● OpenAI API
● Fast!
● Supports prompting
● Post-processing
https://developers.openai.com/api/docs/guides/speech-to-text/
Slide 45
Slide 45 text
https://github.com/ggml-org/whisper.cpp
Slide 46
Slide 46 text
● Runs locally!
● Plain C/C++ implementation
● No python required
● Different models can be used
● CPU/GPU-hungry
https://github.com/ggml-org/whisper.cpp
Slide 47
Slide 47 text
● Large variations in model quality
● --max-context keeps internal consistency high
● --prompt can help with noun recognition
● Prompt too long, and the quality degrades - a balancing act!
Slide 48
Slide 48 text
No content
Slide 49
Slide 49 text
https://github.com/openai/whisper/discussions/194
Slide 50
Slide 50 text
● Sporadic issues with proper nouns
● Challenges with punctuation & run-on sentences
● Occasionally gets a little confused…
https://github.com/openai/whisper/discussions/194
Slide 51
Slide 51 text
Symphony of the butt: https://everythingisshowbiz.com/?episode=323&segment=390151
Slide 52
Slide 52 text
● Transcription breaks on
pause
● Lots of short segments
● Not great for search!
Slide 53
Slide 53 text
No content
Slide 54
Slide 54 text
No content
Slide 55
Slide 55 text
● Transcription breaks on
pause
● Lots of short segments
● Not great for search!
Slide 56
Slide 56 text
● Transcription breaks on
pause
● Lots of short segments
● Not great for search!
Slide 57
Slide 57 text
TRANSCRIBE SERVE
INGEST
A Simple Plan
Slide 58
Slide 58 text
TRANSCRIBE SERVE
REFINE TEXT
INGEST
A Simple Plan
Slide 59
Slide 59 text
INGEST TRANSCRIBE SERVE
REFINE TEXT
A Simple(-ish) Plan
Slide 60
Slide 60 text
What’s in a name?
● Models trained on “mid-Atlantic” accent
● Show features strong regional accents
● Names pronounced differently in each episode
● Post-transcription processing needed
Slide 61
Slide 61 text
No content
Slide 62
Slide 62 text
No content
Slide 63
Slide 63 text
No content
Slide 64
Slide 64 text
No content
Slide 65
Slide 65 text
Actual Conversation
Slide 66
Slide 66 text
Actual Conversation
Original Transcription
Slide 67
Slide 67 text
Responsibility
● Content is AI-generated, but responsibility is ours
● Need guardrails to prevent incorrect publication
● Errors already costing companies
Slide 68
Slide 68 text
Responsibility
● Content is AI-generated, but responsibility is ours
● Need guardrails to prevent incorrect publication
● Errors already costing companies
https://www.theguardian.com/world/2024/feb/16/air-canada-chatbot-lawsuit
Slide 69
Slide 69 text
No content
Slide 70
Slide 70 text
No content
Slide 71
Slide 71 text
No content
Slide 72
Slide 72 text
No content
Slide 73
Slide 73 text
No content
Slide 74
Slide 74 text
No content
Slide 75
Slide 75 text
INGEST TRANSCRIBE SERVE
REFINE TEXT
A Simple(-ish) Plan
Slide 76
Slide 76 text
No content
Slide 77
Slide 77 text
🎉🥳
Slide 78
Slide 78 text
No content
Slide 79
Slide 79 text
No content
Slide 80
Slide 80 text
No content
Slide 81
Slide 81 text
No content
Slide 82
Slide 82 text
INGEST TRANSCRIBE SERVE
REFINE TEXT
A Simple(-ish) Plan
Slide 83
Slide 83 text
INGEST TRANSCRIBE SERVE
REFINE TEXT
SEARCH
SERVE
A Simple(-ish) Plan
Slide 84
Slide 84 text
INGEST TRANSCRIBE SERVE
REFINE TEXT SEARCH
A Plan
Slide 85
Slide 85 text
Semantic Search
● Keyword search asks: “Does this text contain these words?”
● Semantic search asks: “Is this about the same thing?”
● It tolerates synonyms, paraphrasing, and noise
● Under the hood: high-dimensional math
Slide 86
Slide 86 text
Semantic Search
● Keyword search asks: “Does this text contain these words?”
● Semantic search asks: “Is this about the same thing?”
● It tolerates synonyms, paraphrasing, and noise
● Under the hood: high-dimensional math
● ✨✨ Embeddings! ✨✨
Slide 87
Slide 87 text
● Complex data reduced to
simpler information
[159, 837, 528, 162, …]
● The larger the embedding,
the more “meaning” captured
● Similar words or phrases are
close to each other in this
multi-dimensional space
Slide 88
Slide 88 text
● Complex data reduced to
simpler information
[159, 837, 528, 162, …]
● The larger the embedding,
the more “meaning” captured
● Similar words or phrases are
close to each other in this
multi-dimensional space
Slide 89
Slide 89 text
Embeddings
● Postgres extension pgvector
● Generate embedding for each text block
● Store embedding alongside original text
● Generate embedding for each query
● Search by similarity
https://github.com/pgvector/pgvector
My interesting query
Slide 90
Slide 90 text
Embeddings
● Postgres extension pgvector
● Generate embedding for each text block
● Store embedding alongside original text
● Generate embedding for each query
● Search by similarity
https://github.com/pgvector/pgvector
My interesting query
[938, 324, 5, 234.. ]
Slide 91
Slide 91 text
Embeddings
● Postgres extension pgvector
● Generate embedding for each text block
● Store embedding alongside original text
● Generate embedding for each query
● Search by similarity
https://github.com/pgvector/pgvector
My interesting query
[938, 324, 5, 234.. ]
[11, 903, 675, 837.. ]
Slide 92
Slide 92 text
Embeddings
● Postgres extension pgvector
● Generate embedding for each text block
● Store embedding alongside original text
● Generate embedding for each query
● Search by similarity
https://github.com/pgvector/pgvector
My interesting query
[938, 324, 5, 234.. ]
[11, 903, 675, 837.. ]
Slide 93
Slide 93 text
No content
Slide 94
Slide 94 text
Embeddings
● Use local model (Ollama) or API (OpenAI)
● Really basic model is ok! (text-embedding-3-small)
● Makes remote APIs very cheap
Slide 95
Slide 95 text
How cheap exactly?
Slide 96
Slide 96 text
How cheap exactly?
Slide 97
Slide 97 text
No content
Slide 98
Slide 98 text
Observers
● Hook into model lifecycle events
● Automate side-effects cleanly
● Remove glue code from controllers
● Keep business logic cohesive
Slide 99
Slide 99 text
No content
Slide 100
Slide 100 text
No content
Slide 101
Slide 101 text
No content
Slide 102
Slide 102 text
No content
Slide 103
Slide 103 text
Real-world users!
● Tend to use really short queries
● Hard to parse intent!
● How to handle both?
Slide 104
Slide 104 text
Real-world users!
● Tend to use really short queries
● Hard to parse intent!
● How to handle both?
Slide 105
Slide 105 text
Real-world users!
● Tend to use really short queries
● Hard to parse intent!
● How to handle both?
Slide 106
Slide 106 text
No content
Slide 107
Slide 107 text
No content
Slide 108
Slide 108 text
Sub query, avoid recalculation
Slide 109
Slide 109 text
Boost score for exact match
Sub query, avoid recalculation
Slide 110
Slide 110 text
No content
Slide 111
Slide 111 text
No content
Slide 112
Slide 112 text
No content
Slide 113
Slide 113 text
No content
Slide 114
Slide 114 text
● Search for text segments
● Click segment to autoplay
● Share deeplink
● URLs always update!
Slide 115
Slide 115 text
🤨
🤖
Issue Spotted
Slide 116
Slide 116 text
🤨
🤖
🚀
Issue Spotted Magic Fix
Slide 117
Slide 117 text
🤨
🤖
🚀
📦
Issue Spotted Magic Fix
Change Integrated
Slide 118
Slide 118 text
🤨
🤖
🚀
📦
Issue Spotted Magic Fix
Change Integrated
But then…
Slide 119
Slide 119 text
What did I expect?
Slide 120
Slide 120 text
What did I expect?
● Discrete function classes /
jobs
Slide 121
Slide 121 text
What did I expect?
● Discrete function classes /
jobs
● Logical separation
Slide 122
Slide 122 text
What did I expect?
● Discrete function classes /
jobs
● Logical separation
● Testable
Slide 123
Slide 123 text
What did I expect?
● Discrete function classes /
jobs
● Logical separation
● Testable
Slide 124
Slide 124 text
What did I expect?
● Discrete function classes /
jobs
● Logical separation
● Testable
● That’s it!
Slide 125
Slide 125 text
No content
Slide 126
Slide 126 text
No content
Slide 127
Slide 127 text
No content
Slide 128
Slide 128 text
● Side by side of different code
● Highlight “same but different”
approach
● AI solving from 10ft, I’m
approaching this from 10,000ft
Slide 129
Slide 129 text
● Side by side of different code
● Highlight “same but different”
approach
● AI solving from 10ft, I’m
approaching this from 10,000ft
Slide 130
Slide 130 text
● Side by side of different code
● Highlight “same but different”
approach
● AI solving from 10ft, I’m
approaching this from 10,000ft
Slide 131
Slide 131 text
● Side by side of different code
● Highlight “same but different”
approach
● AI solving from 10ft, I’m
approaching this from 10,000ft
Slide 132
Slide 132 text
What’s happening?
● Lots of near-duplicate code
● Huge amount of dead code
● AI is focusing on low-level context alone
Working at 10ft while I’m at 10,000ft
● Need to match the code to my mental model
Slide 133
Slide 133 text
Cognitive debt
https://margaretstorey.com/blog/2026/02/09/cognitive-debt/
Slide 134
Slide 134 text
Cognitive debt
https://margaretstorey.com/blog/2026/02/09/cognitive-debt/
Slide 135
Slide 135 text
Use AI to tidy up AI
● Much tighter feedback loops
● Personas
○ “You are a senior PHP dev…”
● Less vibes, more control
● Closer to pair programming
Slide 136
Slide 136 text
Use AI to tidy up AI
● Much tighter feedback loops
● Personas
○ “You are a senior PHP dev…”
● Less vibes, more control
● Closer to pair programming
Slide 137
Slide 137 text
SERVE
A Plan
INGEST TRANSCRIBE REFINE TEXT SEARCH
Slide 138
Slide 138 text
Workload
Generation
● Heavy GPU load
● But only 2-3x a week
● Admin tooling
Public site
● Effectively read-only
● Updated 2-3x a week
● Strong caching candidate
Slide 139
Slide 139 text
Workload
Generation
● Heavy GPU load
● But only 2-3x a week
● Admin tooling
Public site
● Effectively read-only
● Updated 2-3x a week
● Strong caching candidate
💸💸💸
Slide 140
Slide 140 text
Workload
Generation
● Heavy GPU load
● But only 2-3x a week
● Admin tooling
Public site
● Effectively read-only
● Updated 2-3x a week
● Strong caching candidate
💸💸💸 💸
Slide 141
Slide 141 text
Mac Mini
RSS Feed
(Cheap!) VPS
Slide 142
Slide 142 text
Mac Mini
RSS Feed
(Cheap!) VPS
Slide 143
Slide 143 text
Mac Mini
RSS Feed
(Cheap!) VPS
Slide 144
Slide 144 text
Mac Mini
RSS Feed
(Cheap!) VPS
Slide 145
Slide 145 text
🧑💻
Mac Mini
RSS Feed
(Cheap!) VPS
Slide 146
Slide 146 text
🧑💻
Mac Mini
RSS Feed
(Cheap!) VPS
Slide 147
Slide 147 text
No content
Slide 148
Slide 148 text
No content
Slide 149
Slide 149 text
No content
Slide 150
Slide 150 text
No content
Slide 151
Slide 151 text
No content
Slide 152
Slide 152 text
No content
Slide 153
Slide 153 text
No content
Slide 154
Slide 154 text
No content
Slide 155
Slide 155 text
No content
Slide 156
Slide 156 text
No content
Slide 157
Slide 157 text
No content
Slide 158
Slide 158 text
No content
Slide 159
Slide 159 text
No content
Slide 160
Slide 160 text
No content
Slide 161
Slide 161 text
No content
Slide 162
Slide 162 text
A Plan
INGEST SERVE
TRANSCRIBE REFINE TEXT SEARCH
Slide 163
Slide 163 text
TIDY UP
A Plan
INGEST SERVE
TRANSCRIBE REFINE TEXT SEARCH
Slide 164
Slide 164 text
INGEST TRANSCRIBE
SERVE
REFINE TEXT
SEARCH
Slide 165
Slide 165 text
INGEST TRANSCRIBE
SERVE
REFINE TEXT
SEARCH TIDY UP
Slide 166
Slide 166 text
Late 2025
● Step-change in AI tooling
● Moving from the IDE to CLI
● From “check this code I wrote”
to “I’ll check the code you
wrote”
Slide 167
Slide 167 text
No content
Slide 168
Slide 168 text
No content
Slide 169
Slide 169 text
AI-Assisted Engineering
● Vibe coding optimises for speed
● Engineering optimises for sustainability
Slide 170
Slide 170 text
https://addyosmani.com/blog/ai-assisted-engineering-idea/
AI-Assisted Engineering
● Vibe coding optimises for speed
● Engineering optimises for sustainability
Slide 171
Slide 171 text
https://addyosmani.com/blog/ai-assisted-engineering-idea/
AI-Assisted Engineering
● Vibe coding optimises for speed
● Engineering optimises for sustainability
Slide 172
Slide 172 text
Parallel & Structured
● Sub-agents with defined roles
● Reusable skills
(Laravel clean code, frontend design…)
● Worktrees for isolation
● Inter-agent communication
🤖
🤖 🤖 🤖
The Boss
Design Security Architect
Slide 173
Slide 173 text
From prompting to planning
● Interview with the model first
● Capture decisions in plan.md
● Make constraints explicit
● Break work into checkable steps
Slide 174
Slide 174 text
From prompting to planning
● Interview with the model first
● Capture decisions in plan.md
● Make constraints explicit
● Break work into checkable steps
Slide 175
Slide 175 text
No content
Slide 176
Slide 176 text
No content
Slide 177
Slide 177 text
No content
Slide 178
Slide 178 text
Ralphing
● Models cut corners on long tasks
● Hook in to finished state
● “Did you really finish? Hmmmm???”
● Keep going relentlessly until it’s done
https://github.com/anthropics/claude-code/blob/main/plugins/ralph-wiggum/README.md
Slide 179
Slide 179 text
Cleanup Complete!
● It works (mostly..)
● Documented
● Tested
● Public
https://github.com/conroyp/podcast-transcription
Slide 180
Slide 180 text
Side Projects!
● Abstract away the boring boilerplate
● Dopamine hit of fast feedback
● Keep the momentum going
● Quick to prototype & validate
● From noodling on phone to live app
Slide 181
Slide 181 text
https://www.kidssudoku.com https://www.tictacgoaway.com https://www.jdcaptcha.com
Slide 182
Slide 182 text
https://www.kidssudoku.com https://www.tictacgoaway.com https://www.jdcaptcha.com
Slide 183
Slide 183 text
https://www.kidssudoku.com https://www.tictacgoaway.com https://www.jdcaptcha.com
Slide 184
Slide 184 text
Where does that leave us developers?
● AI is already faster than most mid-level devs
● Productivity up 5-6× → headcount down?
● “Good enough” beats “well-engineered.”
● Why hire juniors when AI does the boilerplate?
Slide 185
Slide 185 text
No content
Slide 186
Slide 186 text
Geoffrey Hinton,
“Godfather of AI”
Slide 187
Slide 187 text
“[I]f you work as a radiologist you are like
Wile E. Coyote in the cartoon. You’re
already over the edge of the cliff, but
you haven’t yet looked down.
Geoffrey Hinton,
“Godfather of AI”
Slide 188
Slide 188 text
“[I]f you work as a radiologist you are like
Wile E. Coyote in the cartoon. You’re
already over the edge of the cliff, but
you haven’t yet looked down.
It’s just completely obvious that in five
years deep learning is going to do better
than radiologists.”
Geoffrey Hinton,
“Godfather of AI”
Slide 189
Slide 189 text
What Happened Next?
● Demand went up
● Salaries increased!
● The nature of the work
changed
Slide 190
Slide 190 text
What Happened Next?
● Demand went up
● Salaries increased!
● The nature of the work
changed
https://worksinprogress.co/issue/the-algorithm-will-see-you-now/
Slide 191
Slide 191 text
What Happened Next?
● Demand went up
● Salaries increased!
● The nature of the work
changed
https://worksinprogress.co/issue/the-algorithm-will-see-you-now/
Slide 192
Slide 192 text
Who gets to build?
● Doctors can ship apps
● Lawyers can prototype tools
● Contractors can automate their workflows
● Founders can validate ideas without devs (?)
Slide 193
Slide 193 text
We’ve been here before
● Assembly → C
● C → PHP
● PHP → Low-/no-code tools
● Manual infra → DevOps
● On-prem → Cloud
Slide 194
Slide 194 text
We’ve been here before
● Assembly → C
● C → PHP
● PHP → Low-/no-code tools
● Manual infra → DevOps
● On-prem → Cloud
https://addyosmani.com/blog/the-efficiency-paradox/
“Every time we’ve made it
easier to write software, we’ve
ended up writing exponentially
more of it.”
Slide 195
Slide 195 text
What changes for us?
● Execution is getting cheaper
● Prototypes are cheaper
● Boilerplate is getting cheaper
● Judgment is not
Slide 196
Slide 196 text
What changes for us?
● Execution is getting cheaper
● Prototypes are cheaper
● Boilerplate is getting cheaper
● Judgment is not
https://addyosmani.com/blog/the-efficiency-paradox/
The real question is whether
we’re prepared for a world
where the bottleneck shifts
from “can we build this?” to
“should we build this?”
Slide 197
Slide 197 text
No content
Slide 198
Slide 198 text
Thank you!
Conroyp.com
@conroyp
[email protected]
🌍
🌐
📧
https://joind.in/event/php-uk-conference-2026/everything-is-showbiz-lessons-from-a-php--ai-side-projects