Projects
Courses, sites, and project plans in one place.
Courses
Hosted courses across ifkash.dev, dotslasha.me, and Pages
-
Machine learning
A graduate-level course from least squares to mixtures of experts, with a check runner beside every lesson.
ml.ifkash.dev -
LLM under the hood
Build every layer of a large language model from scratch.
llm-under-the-hood.pages.dev -
vLLM course
An interactive course on vLLM and how it serves large language models.
vllm-course.pages.dev -
Learn inference
An inference-engineering course with a lab runner that executes on a GPU.
qwen.ifkash.dev -
Statistics
Twenty-eight chapters from class 12 to the research frontier, each written for four levels of reader.
stats.dotslasha.me
Static sites
Single-page sites deployed through Dokploy and Garage
-
Teaching a language model to speak
From raw waveforms to a speaking Gemma 4, through SoundStream, AudioLM, and a grafted speech head.
dokployteach-llm-to-speak -
Signal to symbol
A learning path from training language models to voice, audio representations, and streaming generation.
dokployrj-ashok-knowledge-gap -
Cursor plan usage
An infographic of 149 logged Cursor requests: spend, tokens, cache hits, and usage by model.
garagecursor-start-usage -
Mochi demo briefing
A meeting briefing and live demo guide for Mochi, an open-source AI agent.
garagemochi-demo-eric -
Sherry Moore call prep
Framing, talking points, and next steps for a first mentorship call.
garagesherry-prep
Projects
Weekend build plans, each with a spec for a coding agent
-
canopy
Linear attention broke tree speculation, and you write the one-solve fix
plan33 min read -
clatter
Hear what any 3D object sounds like, live in your browser
plan19 min read -
dreamcart
A video game with no game code, dreamed by a $4 chip
plan18 min read -
GeoGuessr RL
Post-train an open-weights VLM with GRPO in one weekend
plan22 min read -
mindseye
A robot that plans in its imagination, live in your browser tab
plan16 min read -
monolith
One CUDA kernel runs a whole language model on a gaming GPU, and you can watch every SM
plan30 min read -
rewind
An AI weather model that predicts yesterday, live in your browser
plan23 min read -
sprout
A knob learned on one fish reshapes creatures it has never seen, live in your browser
plan28 min read -
static
A network learns to read handwriting from TV static, live in your browser
plan29 min read -
tally
Catch cuBLAS dropping eight numbers from a sum, then redraw how it adds
plan30 min read -
thirteen
Double-precision math from 4-bit numbers, about 8× faster than the GPU's own FP64
plan34 min read
Archived 31 items from the old projects board
Todo
- WisprClone
- CaveCode
- LLM to VLM
- Kimi K3 implementation
- day officework
- learning how to host VPN — ask enderboi to teach
- complete enderboi's course (the website)
In Progress
- c2m
- muchi
- Find cream people in Google
day officework · 21
- [P0] Review Sherry recording
- Humanize featuregaps doc
- Prepare dataset for c2m
- Complete Alankrit resources study (MMM + related)
- Journal 1/3 — Update existing 'Mochi's Day' note into a manager-meets journal
- Journal 2/3 — Create Talentsprint journal to track content they teach
- Journal 3/3 — Create purpose journal from their template (work log) + explore automation
- Figure out SSH for local development
- Set up Obsidian for resources/links (like Mirchiea shared)
- [P0] Finalize competitors analysis doc + send the mail
- Search internal resources for Mochi
- Text Gazi
- Check FEI docs
- Search for other 20% projects
- Check out and complete tasks in Busy Beaver
- Gemma — watch Fireship video, read up RLHF at Google, look into Gemma core team (note: core team is in UK — advantage for me)
- Ask Mirchie how to get SSH/terminal access locally
- Ask Naveen/Anuran for the update link
- [P0] Study the paper Matt sent (check what he sent)
- Look into repo 'OneTwo' (Mirchie's team uses it, seems opensource) — see if I can contribute
- research TPU stuff at Google — who is working on what