Inference Engineering
Philip Kiely · Baseten
“The more constraints you can introduce into your inference system, the better performance you will achieve.” The Golden Rule of Inference.
keypaa/EFREI Paris/Paris, France
KeylhanPaumard–André
Tinkering with AI from a Linux box in Paris.
I'm Keylhan, an engineering student at EFREI Paris (just finished my first year, année préparatoire), French, daily-driving Linux with Windows on the side for school. I read a lot, lurk on X more than I post, and tinker with AI the way the hardware allows: no GPUs, so I go where the interesting work doesn't need them, going after reverse-engineering how tools are built, local-first infra, clever scripts, and the odd borderline-legal extraction.
↓scroll
now
A snapshot of what's current.
readingTalking to Strangers, Malcolm Gladwell
buildingNestClockOS, bilingual voice smart clock
untanglingMinified symbols in the Claude Code binary
based inParis, France
schoolEFREI Paris, year 1 / 5
reading log
A commonplace book of what I'm reading.
3 read · 2 on the queue
recently read
read
The Pragmatic Programmer
Andrew Hunt & David Thomas · 20th Anniversary Edition
Good design is always easier to change, and great code starts with taking ownership.
Pourquoi les classiques
Italo Calvino
“A classic is a book that has never finished saying what it has to say.” Fourteen definitions of a classic, and no reason better than this one.
on the queue
Talking to Strangers
Malcolm Gladwell
“The right way to talk to strangers is with caution and humility.” The closing thesis of a book about how badly we misread people we don't know.
The O'Reilly catalog
ongoing, broad goal
Read as much as I can from O'Reilly over the coming months / years.
workbench
Things I've built, broken, and reverse-engineered.
6 entries · no GPUs harmed
Claude Web sandbox: rebuilt from scratch
liveReverse-engineered how Claude's web sandbox works and rebuilt it from the ground up.
Curiosity-driven teardown. Figured out the moving parts, then re-implemented them myself to actually understand the design.
reverse-engineeringfrom-scratch
github.com/keypaa/fuseboxllama.cpp easy-install for cloud instances
liveA script that skips the 20-minute llama.cpp compile on 4-core Colab sessions.
I burned a lot of Colab sessions recompiling llama.cpp every time: 4 cores, 20+ minutes each. So I scripted the painful part away. Now a fresh cloud instance is usable in minutes.
toolinginfralocal-first
github.com/keypaa/llamaupClaude Code binary: reverse engineered
in progressExtracted the .exe, rebuilt the source tree from the extracted contents, now I can recompile with my own patches.
A bit borderline, legally, but a great exercise. Currently working on recovering meaningful function names from the minified output, so LLMs can actually navigate the codebase instead of drowning in garbage symbols.
reverse-engineeringtooling
github.com/keypaa/0xClaudezenno, a Claude Code plugin
liveSecurity guards, disciplined multi-agent orchestration, and persistent memory, for Claude Code and opencode.
Zenno watches what agents do and keeps them honest: four Shield guards block secret leaks, confidential-file reads, risky installs, and self-modification of Zenno itself; hard caps stop runaway subagent fan-out; a memory layer carries project knowledge across sessions; and a doctor, cost tracker, and trace exporter keep the whole thing observable. I run it on my own machine every day.
plugintoolingagents
github.com/keypaa/zennoNestClockOS: bilingual voice smart clock
in progressA voice-controlled clock that runs its command routing on-device: mic → Voxtral STT → a locally fine-tuned Laya router, cloud LLM only as fallback. Speaks French and English, including mixed.
The interesting part wasn't the app, it was the measurement. Zero-shot routing scored 91% in English but 62% in French, so I generated 2,500 labeled commands, fine-tuned Laya's multilingual checkpoint with RLCD on a Colab T4, and took French to 93%, then spent two more rounds chasing the confident mistakes that remained, and learned when to stop training and fix safety in plain code instead.
voiceon-devicefine-tuning
more soon
Homelab stack
live13 compose projects on an OptiPlex-3070, all behind Traefik, plus a Hermes agent that does the daily driving.
The substrate under most of my experiments: photos, media, bookmarks, DNS, and the dashboards that tell me the box is still alive. Intel UHD 630, no discrete GPU, everything on one 1TB drive and a 256GB NVMe. Watchtower keeps the images fresh, which is the least glamorous part of owning any of this.
- AI
- open-webui
- Media
- immich · jellyfin · youtube-dl
- Reading
- karakeep
- Network
- traefik · adguard · duckdns · portainer
- Ops
- glance · watchtower · speedtest
infraself-hostedagents
more soon
agent usage
I run a lot of agents. Here are the receipts.
Aug 1 – Sep 28, 2026 across opencode + Claude Code, 283 sessions, 5.04B tokens processed, $0.00 spent. Being a free-tier hacker has its perks. Sessions count the ones I started myself, with subagent fan-out left out.
sessions
283
37 active days
messages
26,396
to/from agents
tokens
5.04B
processed
total cost
$0.00
free tier, preview, or via opencode
from cache
92%
of all tokens
sessions / day
peak 59 · 08-23Aug 01Sep 28
top models
by sessionsmuse-spark-1.365 sess · 1.92B
muse-spark-1.239 sess · 1.35B
x-preview-f33 sess · 79.5M
deepseek-v4-flash20 sess · 71.7M
top projects
by tokensVision-Adapter1.69B · 30 sess
claude-desktop-omarchy1.08B · 23 sess
0xClaude749.6M · 27 sess
dsh-jspace-workspace131.2M · 8 sess
exported from opencode + Claude Code analytics · Aug 01, 2026 → Sep 28, 2026 · updated Sep 28, 2026. cache reads account for 92% of all tokens: the prompt cache does the heavy lifting. analytics tools estimate $1,891 for the same usage, since they price free models at list rate.
elsewhere
Find me, or say hi.
I lurk more than I post, but I read everything and I'm always up for a conversation about AI tooling, reverse-engineering, or a good book.