Projects Playground Blog Timeline Get in touch
Writing & Experiments

Notebooks, deep dives
& things I figured out.

Where I document experiments, share failures, and write about building AI systems that actually work in production.

Filter
📌 Pinned
BosWeigh — weighing a cow with one photo, and why it's harder than it looks

Monocular depth estimation + five anatomical keypoints + a 150-year-old livestock formula. The pipeline runs end-to-end. The numbers say I've still got work to do. An honest, mid-flight write-up of a research project — including the negative R² I'm trying to fix.

Apr 20, 2026 · 9 min read · ● New
Computer Vision DepthPro Experiment WIP
Read post →
Notebook preview
Final_Weight_estimator.ipynb
[1]
depth = DepthPro(image) # metric depth, m
[2]
kpts = yolo.predict(image).keypoints
## integrate 3D arc along girth line
[3]
arc = np.linalg.norm(np.diff(pts3d, 0), axis=1).sum()
[4]
w = (length_in * girth_in**2) / 660
Out
Estimated Weight (kg): 258.57
Out
[ 3D arc length chart ]
All posts — 14 total
Sort: newest first ↓
Apr 21, 2026·12 min read·● New
GPT-2 from scratch — understanding attention by writing it

Rebuilt the 124M GPT-2 stack line-by-line in PyTorch — multi-head attention, GELU, transformer blocks — the cleanest way I've found to actually understand attention. With every important code snippet and the six things that only landed once I wrote it.

LLMPyTorchFrom ScratchDeep Dive
gpt2.ipynb
[1]
class MultiHeadAttention(nn.Module):
## causal mask + scaled dot-product
[2]
attn = softmax(Q @ K.T / sqrt(d_k))
[3]
model = GPTModel(GPT_CONFIG_124M)
Out
params: 124M ✓
Mar 14, 2026·12 min read
Fine-tuning LLaMA 3 on agricultural domain data — what 10 failed runs taught me

I tried to get a 7B model to understand cattle health, crop cycles, and Kannada farming terminology. Here's exactly what went wrong and what eventually worked — with full loss curves and prompting strategies.

LLMFine-tuningExperiment
llama_agri_finetune.ipynb
[1]
model = load_pretrained("llama-3-8b")
[2]
trainer.train(resume_from_checkpoint=True)
Out
Epoch 3/5 — val_loss: 0.89 → 0.42 ✓
Feb 28, 2026·5 min read
When YOLO fails in the real world: edge cases in production CV pipelines

Low-light, motion blur, and occlusion — the failure modes benchmarks never cover, and how I found them in production.

Computer VisionDeep Dive
yolo_edge_cases.ipynb
[1]
yolo.predict(img, conf=0.25)
Out
0 detections — night-time failure
[2]
analyze_failures(results_df)
Jan 15, 2026·7 min read
Building a RAG pipeline that doesn't hallucinate: lessons from GauSwastha

How we combined retrieval with structured livestock data to get reliable, explainable outputs from an LLM.

RAGLLMExperiment
rag_gauswastha.ipynb
[1]
retriever = FAISSRetriever(docs)
Out
Top-3 retrieved: [doc_1, doc_8, doc_12]
[2]
chain.run(query, context)
Dec 4, 2025·4 min read
Markowitz optimization with live market data in Python

A step-by-step walkthrough of the mean-variance optimizer I built — constraints, edge cases, and the math behind it.

Financial MLTutorial
markowitz_optimizer.ipynb
[1]
weights = optimize(returns, cov)
Out
Sharpe: 1.42 | Sortino: 1.87 ✓
[2]
plot_efficient_frontier()