┌──────────────────────────────────────────────────────────────────────────────┐
│ │
│ ███████╗███████╗██╗ ██╗██╗ ██╗ │
│ ██╔════╝██╔════╝██║ ██║╚██╗██╔╝ │
│ █████╗ █████╗ ██║ ██║ ╚███╔╝ │
│ ██╔══╝ ██╔══╝ ██║ ██║ ██╔██╗ │
│ ██║ ███████╗███████╗██║██╔╝ ██╗ │
│ ╚═╝ ╚══════╝╚══════╝╚═╝╚═╝ ╚═╝ │
│ │
│ Kenya → Germany | Deriving gradients, not importing them │
│ │
└──────────────────────────────────────────────────────────────────────────────┘
class Felix:
"""
Software engineer who believes understanding beats memorization.
What I cannot create, I do not understand.
"""
location = "Germany"
origin = "Kenya"
timezone = "CET (UTC+1)"
by_day = "InfoSec Compliance & Data Governance"
by_night = "Building neural networks from scratch"
philosophy = "Ask 'why' until the abstractions dissolve"
currently_reading = ["Kafka", "Orwell", "McCarthy"]
# Writers who make every word earn its place
def approach(self, problem):
while not understood_from_first_principles(problem):
ask_why()
derive_it_myself()
validate_against_known_implementation()
return deep_understanding+ Derived backpropagation from limit definitions, validated against TensorFlow
+ 5-6 hours per Karpathy video — implementing, not just watching
+ Built 25+ production workflows in enterprise data governance
+ Same rigor whether the deadline is compliance audit or gradient descent
- Tutorial following
- "It works, ship it"
- Surface-level understanding● learning.service - Neural Networks: Zero to Hero
Loaded: loaded (/karpathy/nn-zero-to-hero; enabled)
Active: active (running) since 2024
Docs: https://youtube.com/@AndrejKarpathy
Process: implementing from memory after each video
Memory: matrix multiplications and chain rules
MISSION PROGRESS
════════════════════════════════════════════════════════════════════════
[■■■■■■■■■■■■■■■■■■■■■■■■■■■■░░░░░░░░░░░░░░░░░░░░░░] 57% ████▌
Part 1 [████] DONE Micrograd ─────────── Autograd from first principles
Part 2 [████] DONE Makemore ──────────── Bigram language model
Part 3 [████] DONE Makemore 2 ────────── MLP with embeddings
Part 4 [████] DONE Makemore 3 ────────── Activations & gradients
Part 5 [▓▓░░] ACTIVE BatchNorm ─────────── Normalizing activations
Part 6 [░░░░] QUEUED Wavenet ───────────── Dilated convolutions
Part 7 [░░░░] QUEUED GPT ───────────────── The endgame
════════════════════════════════════════════════════════════════════════
What "DONE" actually means
Not: watched the video
Not: copied the code
Actually:
├── Implemented from scratch without looking
├── Derived the math by hand (pen + paper)
├── Validated against production frameworks
├── Can explain it to someone else
└── Can rebuild it tomorrow from memory
[2026-01-25 01:30:42] Solved: Matrix transformation T⁻¹AS (Deep-ML)
[2026-01-24 23:15:00] Insight: Invertibility requires det ≠ 0 — geometry clicked
[2026-01-24 20:00:00] Reading: Kafka — paranoia as structure
[2026-01-23 19:30:00] Aha: Learning rate decay = coarse-to-fine search
[2026-01-22 21:00:00] Orwell: "Never use a long word where a short one will do"
Deep-ML Daily Practice — building fluency, one problem at a time
MATHEMATICAL FOUNDATION
════════════════════════════════════════════════════════════════════════
Linear Algebra MIT 18.06 (Strang) ████████████████████░ 7.5/8
Not memorized — understood geometrically
Column spaces, projections, four fundamental subspaces
Calculus Stewart + Thompson ████████████████████ Complete
Derivatives from limits, chain rule derived not memorized
Probability MIT 6.041 ████████████████░░░░ 70%
Just-in-time: learn when needed, not exhaustively
════════════════════════════════════════════════════════════════════════
Foundation before frameworks. Always.
════════════════════════════════════════════════════════════════════════
drwxr-xr-x ml_stack/
├── numpy ████████████████████ The foundation — I think in arrays
├── pytorch ████████████████░░░░ Building up, not importing blindly
├── pandas ████████████████░░░░ Data wrangling
└── jupyter ████████████████░░░░ Interactive exploration
drwxr-xr-x languages/
├── python ████████████████████ Primary weapon
├── javascript ████████████████░░░░ Full-stack when needed
├── groovy ████████████████░░░░ Collibra workflows
└── sql ████████████████░░░░ Data extraction
drwxr-xr-x enterprise/
├── collibra ████████████████████ 25+ production workflows
├── sap_datasphere████████████████░░░░ Data warehousing
├── databricks ████████████████░░░░ Big data
└── compliance ████████████████████ ISO, SOC2, GDPR
[philosophy]
core = "What I cannot create, I do not understand"
approach = "first_principles"
questions = "why > how"
[method]
step_1 = "Encounter the problem"
step_2 = "Resist the urge to Google"
step_3 = "Derive it myself (pen + paper)"
step_4 = "Implement from understanding"
step_5 = "Validate against known implementations"
step_6 = "Rebuild from memory tomorrow"
[anti-patterns]
reject = ["tutorial-following", "copy-paste", "it-works-ship-it"]
embrace = ["confusion", "being stuck", "productive struggle"]
[timeline]
patience = "multi-year"
motto = "Precept upon precept, line upon line"╔═══════════════════════════════════════════════════════════════════════════╗
║ ║
║ Felix Onyango ║
║ ───────────────────────────────────────────────────────────────────── ║
║ Email: jalochglitch@gmail.com ║
║ GitHub: github.com/Jaloch-glitch ║
║ Location: Germany ║
║ Timezone: CET (UTC+1) ║
║ ║
║ Open to: ML collaborations • Hard problems • First-principles work ║
║ ║
╚═══════════════════════════════════════════════════════════════════════════╝
┌─────────────────────────────────────────────────────────────────────────┐
│ "What I cannot create, I do not understand." — Feynman │
└─────────────────────────────────────────────────────────────────────────┘