Capstone: My College AI Portfolio 🎓

a) Serve → Fine-tune → Retrieve

b) Fine-tune the base model on domain data (QLoRA) → wrap inference in a hybrid RAG pipeline so the model has up-to-date facts at query time → serve at scale with vLLM + Kubernetes for cost-efficient throughput. Each stage builds on the previous: fine-tuning gives base capability, RAG provides freshness, vLLM provides scale.

c) Retrieve → Serve → Fine-tune

d) The order doesn't matter as long as all three are used

a) Always use the largest batch size your GPU memory allows

b) Distributed training requires careful data partitioning — DistributedSampler ensures each GPU sees a different shard each epoch with no overlap; NeighborLoader samples each node's neighbourhood independently to fit graph batches in memory. In both cases, naive batching either duplicates work or runs out of memory; the right sampler is the unlock.

c) Data parallelism is always better than model parallelism

d) Mini-batching is only required when training on multiple GPUs

a) Always charge premium prices for AI-generated content

b) Generative AI requires consent + provenance — synthesising someone's likeness or voice without permission violates Indian DPDPA 2023, IT Rules 2021/2023, and personality rights case law (Anil Kapoor v. Simply Life India). Watermarking and refusing unconsented clones are both expressions of "you do not synthesise or simulate a person without their consent and clear labelling."

c) Open-source models are safer than commercial ones

d) Watermarks reduce model accuracy and should only be added on request

a) Two towers always achieve higher accuracy than single models

b) At serving time, you can precompute and FAISS-index every item embedding once; for each query you only encode the user vector and do a fast nearest-neighbour search. A single combined model would require evaluating one network per (user, item) pair — billions of forward passes per query batch, which does not scale.

c) Two towers can be trained on different programming languages

d) FAISS only supports two-tower architectures

a) AI features always reduce engineering cost

b) Decisions about AI infrastructure and product launches require quantitative reasoning before commitment — Rohit calculated the crossover point (~5K DAU) where self-hosting wins; Devansh calculated the sample size needed for statistical significance. Without these calculations, both would have shipped on intuition and lost either money or evidence.

c) Self-hosting is always cheaper than using APIs

d) A/B tests should be skipped to ship faster

a) Always use the median forecast — outliers should be ignored

b) The right quantile of a forecast distribution depends on the asymmetric cost of being wrong in each direction. Stockout of mithai loses customers (high cost) so order optimistically; overstock of packaging ties up cash (lower cost) so order conservatively. Point forecasts ignore this asymmetry; quantile forecasts make it explicit.

c) Confidence intervals are only for academic papers, not production systems

d) 80% is always the right confidence level for inventory decisions

a) Every problem in AI can be solved with a larger neural network

b) Framing the problem correctly matters more than the model architecture chosen. Treating fraud as tabular missed the relational structure; treating voice synthesis as a tech problem missed the consent failure mode; treating AI features as model problems missed the rollout-and-measurement failure mode. Wrong framing × good model = wrong answer.

c) AI engineers should specialise in only one domain

d) Indian AI problems require different algorithms than global ones

a) Reranking is required by the OpenAI API specification

b) Initial retrieval optimises for recall over a huge candidate pool (billions of items, must be fast); reranking optimises for precision and diversity over a small candidate set (top-100, can be slow). Combining them collapses both budgets — fast retrieval over billions + accurate scoring over hundreds is how production search and recommendations actually scale.

c) Reranking improves model accuracy by 90% in all cases

d) Reranking is only used for English-language results

a) AI projects should target maximum revenue potential first

b) The best AI projects start from a real, named human in your immediate community whose specific problem you understand deeply. Distance-from-the-problem is the leading indicator of irrelevant AI work — every memorable project in this course began with someone the builder personally knew, not a market opportunity from a slide deck.

c) Indian AI must focus on rural problems exclusively

d) AI projects should be entirely funded by personal savings

a) Owning expensive personal hardware (multi-GPU rigs)

b) Shipping working artefacts publicly. Aishwarya pushed her LoRA adapter to Hugging Face. Tara released her Warli posters with credits. Rohit open-sourced his vLLM Helm chart. Ananya pinned 3 flagships on GitHub. They all turned learning into shipped, dated, public work — which is the single most reliable signal college admissions and AI employers have ever had.

c) Coming from English-medium schools in metro cities

d) Having parents already in the AI industry