┌──────────────────────────────────────────────────────────────────────────────┐
│ │
│ ███████╗███████╗██╗ ██╗██╗ ██╗ │
│ ██╔════╝██╔════╝██║ ██║╚██╗██╔╝ │
│ █████╗ █████╗ ██║ ██║ ╚███╔╝ │
│ ██╔══╝ ██╔══╝ ██║ ██║ ██╔██╗ │
│ ██║ ███████╗███████╗██║██╔╝ ██╗ │
│ ╚═╝ ╚══════╝╚══════╝╚═╝╚═╝ ╚═╝ │
│ │
│ 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
Memory: matrix multiplications and chain rules
MISSION PROGRESS
════════════════════════════════════════════════════════════════════════
[■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■░░░░░░░░░░] 78% ████████▌
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 [████] DONE BatchNorm ─────────── Normalizing activations
Part 6 [████] DONE GPT ───────────────── 10.79M param transformer from scratch
Part 7 [▓▓░░] ACTIVE Advanced ──────────── Scaling and production patterns
════════════════════════════════════════════════════════════════════════
What "DONE" actually means
Not: watched the video
Not: copied the code
Actually:
├── Implemented from scratch
├── Derived the math by hand (pen + paper)
├── Validated against production frameworks
└── Can explain it to someone else
[2026-07-14] Built: GPT from scratch — 10.79M params, trained 50 min on Shakespeare
[2026-04-20] Built: Equipment failure LSTM — ablation study, class imbalance analysis
[2026-03-20] Built: CIFAR-10 CNN — conv feature extraction, 72% on 10 classes
[2026-03-17] Built: House price regression — Random Forest, R²=0.81 on California housing
[2026-03-17] Built: Sentiment LSTM — full NLP pipeline from tokenizer to inference
[2026-03-03] Built: MNIST feedforward — 98.5% accuracy, zero convolutions
[2026-03-02] Built: Iris classifier — first production-pattern pipeline
Deep-ML Daily Practice — building fluency, one problem at a time
ML MODEL PORTFOLIO ─ 14 models built
════════════════════════════════════════════════════════════════════════
CLASSIFICATION
──────────────
01 iris_classifier RandomForest 93.3% accuracy
02 mnist_classification Feedforward NN 98.5% accuracy
06 image_classification_cnn CNN (CIFAR-10) 72.5% accuracy
07 transfer_learning ResNet18 fine-tuned 40.6% accuracy
08 object_detection_R-CNN Faster R-CNN COCO pre-trained
10 image_classification_v2 CNN (Fashion MNIST) 91.1% accuracy
11 xray_pneumonia CNN binary 77.2% acc, F1 0.84
REGRESSION
──────────
04 house_price_prediction RandomForest R² = 0.81
05 stock_price_prediction LSTM time-series RMSE $77.93
NLP / SEQUENCE
──────────────
03 sentiment_analysis_imdb LSTM custom tokenizer
14 gpt_from_scratch GPT decoder 10.79M params
ANOMALY / TIME-SERIES
──────────────────────
09 credit_card_fraud MLP + class weights AUC 0.962, 82.4% recall
12 equipment_failure LSTM + ablation 98% acc, F1 analysis
13 log_anomaly_detection RNN autoencoder reconstruction error
════════════════════════════════════════════════════════════════════════
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 ██████████████████░░ Feedforward, CNN, RNN, LSTM, GPT
├── 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
★ gpt_from_scratch 10.79M param GPT decoder — attention derived from scratch
github.com/Jaloch-glitch/gpt_from_scratch
★ equipment_failure_prediction LSTM on sensor time-series + ablation study
github.com/Jaloch-glitch/equipment_failure_prediction
★ image_classification_cnn CIFAR-10 CNN — conv feature extraction, 72.5% on 10 classes
github.com/Jaloch-glitch/image_classification_cnn
★ micrograd Autograd engine from first principles
github.com/Jaloch-glitch/micrograd
[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"
[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 │
└─────────────────────────────────────────────────────────────────────────┘