The Bake Off
Behind every project is a story of struggle, iteration, and discovery. Here's mine — raw and unfiltered.
Where the real
learning happened
The Bake Off is my documentation of the messy, non-linear reality of building projects as a student — the dead ends, the breakthroughs, and the moments that changed how I think about engineering.
The Challenge
Our goal was ambitious: build a system that could identify Indian dishes from photos and recommend Ayurvedic dietary advice based on the user's dosha (Vata, Pitta, or Kapha). The immediate blocker? Accurate food recognition in the wild is genuinely hard — especially for visually similar dishes.
I owned the entire image processing pipeline. We started with raw pixel data, but that approach quickly hit a wall. The model was memorizing, not learning. That's when I turned to HOG (Histogram of Oriented Gradients) — a feature extraction technique that captures shape and texture rather than raw color.
The Process
I ran 11 progressive Random Forest experiments — tweaking tree count, depth, feature subsets, and even reformulating the problem itself. The jump from 37.89% multi-class accuracy to 45.16% came simply from reframing the task as a binary problem. But the real breakthrough was switching to TF-IDF text features derived from ingredient descriptions, which shot accuracy to 75.42%.
That single discovery rewired how I think about data representation. The right features beat the right algorithm, almost every time.
"The most valuable thing I learned wasn't scikit-learn syntax or model evaluation metrics. It was the mindset shift: when your model underperforms, look at your data representation before you blame the algorithm."
Key Metrics
- Algorithms Tested7
- Total Trials26+
- RF Experiments11
- Baseline Accuracy37.89%
- Binary Reformulation45.16%
- Final Accuracy75.42%
- Feature Reduction12×
SDG Alignment
Aligned with UN SDG 2 (Zero Hunger) and SDG 3 (Good Health) — promoting culturally-aware nutritional guidance at scale.
The Context
Our team of six tackled a government-issued problem statement around digital infrastructure and data accessibility. The catch: we had six people with different skill levels, different preferred stacks, and exactly 36 hours to produce something coherent and functional.
Hour one was spent arguing about the tech stack. Hour two was spent agreeing on a shared architecture. Hours three through thirty-six were a blur of caffeine, commits, and compromise.
What I Learned
I handled the backend API layer and spent most of the night wrestling with data schema decisions that kept cascading into frontend breakages. The biggest realization: in a hackathon, good-enough architecture deployed is better than perfect architecture planned.
I also learned how to communicate technical decisions under pressure — not just to teammates but to judges who needed the story, not just the code.
"By hour 30, none of us remembered why we'd made half the architectural decisions we had. That's when I started writing micro-comments in the code — not for the judges, but for my own future self."
Stats
- Team Size6 members
- Duration36 hours
- Git Commits87
- Merge ConflictsToo many
- Coffee Cups∞
- Demo StatusShipped ✓
The Starting Point
I started with Andrew Ng's course on Coursera. By week three, I understood linear regression. By week five, I could implement it from scratch. But something felt off — I was getting better at following along, not at actually solving problems.
The shift happened when I forced myself to close all tutorials and build the KhanPaan project from scratch using only documentation. That's when confusion became competence.
The Real Education
I discovered that ML isn't about memorizing algorithms. It's about understanding what questions your data can answer — and which ones it can't. Feature engineering isn't a trick; it's the whole game. Model selection comes last, not first.
I'm still learning. But now I know how to learn it — through projects, not passive watching.
"Tutorial hell is real. The moment I accepted confusion as part of the process — and stopped expecting instant understanding — everything started to click faster."
Learning Stack
- Courses Completed8+
- Kaggle Notebooks12+
- Models Implemented7+
- Papers Read5+
- Projects Shipped3