Capstone: My AI Research Project 🎓

a) Deploy → Monitor → Optimise

b) Optimise (build efficiently) → Deploy (containerise and ship) → Monitor (detect drift in production) — the pipeline flows from local development to serving to continuous maintenance. Skipping any stage creates technical debt: skipping optimisation creates slow serving, skipping deployment creates "works on my machine" failures, skipping monitoring creates silent degradation.

c) Monitor → Optimise → Deploy

d) The order depends on team size and project methodology

a) Avoid matrix multiplication when possible — it is the bottleneck in all neural network computations

b) Every equation in a paper or textbook is directly implementable as NumPy code — dz = (ŷ - y) maps to an assignment, Attention(Q,K,V) = softmax(QKᵀ/√d_k)V maps to four lines. Math and code are the same language. This principle transforms research papers from intimidating documents into executable programs.

c) NumPy's automatic differentiation eliminates the need to derive gradients manually

d) GPU-based NumPy arrays can run backpropagation 1000x faster than CPU

a) All AI systems require a rate limiter to prevent API abuse

b) Never trust user-controlled input for code execution or instruction override. Deepika found that passing user input to eval() is a critical vulnerability — an attacker can execute OS commands. Meghna found that user messages can override system instructions via prompt injection. Both are instances of "user input should never gain the authority of system instructions."

c) Logging all user interactions is required for both security and legal compliance

d) AI systems should be restarted daily to clear accumulated malicious inputs

a) PPO is always better than Q-learning — it should be used for all RL problems

b) LunarLander has an 8-dimensional continuous state space — tabular Q-learning requires a discrete, enumerable state space. The number of possible states in a continuous space is infinite, making a Q-table impossible. DQN/PPO approximate the Q-function with a neural network, enabling generalisation across the continuous state space that tabular methods cannot handle.

c) CartPole is a simpler game — harder games always require policy gradient methods

d) PPO trains faster on multi-core CPUs, making it the practical choice for games with complex graphics

a) Aryan's app is multi-modal; Deepika's app is text-only

b) Aryan's app is a fixed pipeline (audio → text → image+text → text) where each step is predetermined. Deepika's agent is an autonomous loop — the model decides at runtime which tools to call, in what order, and when to stop. Fixed pipelines are more predictable and easier to audit; agents are more flexible but introduce emergent behaviour that requires careful safety design.

c) LangGraph supports multi-modal inputs; OpenAI tool calling is text-only

d) Aryan's app uses open-source models; Deepika's app is locked to OpenAI's API

a) Technical debugging — finding and fixing errors in model code

b) Risk management — systematically identifying threats to the safe and sustainable operation of an AI system. Data drift threatens model accuracy; legal non-compliance threatens the organisation; poor unit economics threaten viability. Building AI products requires constant risk identification and mitigation across technical, legal, and business dimensions simultaneously.

c) Data science — analysing datasets to extract actionable insights

d) Research — applying academic methods to real-world AI deployment problems

a) Performance optimisation — always prioritise speed over code clarity

b) Defensive engineering — write code that fails loudly and predictably rather than silently. BatchNorm prevents training instability before it can cause silent divergence; type hints + dataclasses surface bugs at definition time rather than at runtime; Redis caching fails predictably with a cache miss (not a system crash). Good engineering is designing your system to reveal problems early and clearly.

c) Abstraction — hiding implementation details behind clean interfaces

d) Modularity — splitting large functions into smaller single-purpose units

a) Gross margin improvement — reducing per-unit cost to increase profit on each sale

b) Distribution leverage — using an existing trusted intermediary (FPO) that already has the relationship with target customers dramatically reduces acquisition cost without changing the product or its price. This is why B2B2C distribution structures dominate in markets where the end customer (farmer) is expensive to reach individually. Unit economics improve by 10–20x with no technical changes.

c) Retention optimisation — FPOs create lock-in that increases average subscription duration

d) Pricing power — FPOs allow higher subscription prices due to institutional trust

a) Government compliance — following mandatory regulations before product launch

b) Human-centred AI — designing AI systems that serve real human needs with dignity and safety, rather than deploying technology for its own sake. Aryan started with his grandmother's real need; Meghna started with student safety; Rahul started with farmer pain. The most impactful AI applications emerge from deeply understanding the humans they serve before writing a single line of code.

c) User research methodology — applying ethnographic techniques to product development

d) Social entrepreneurship — choosing mission-driven applications over commercially viable ones

a) Access to expensive hardware — all successful AI builders need powerful GPUs from the start

b) The discipline to implement, not just read. Rohan ran the profiler. Ananya derived the gradients by hand. Vikram trained the CartPole agent. Every character moved from "I understand this concept" to "I have running code that proves I understand it." Implementation reveals the gaps in your understanding that reading never can. This is the skill that separates engineers from enthusiasts.

c) Strong mathematical background — linear algebra and calculus are prerequisites for all AI work

d) Connections to established AI researchers and access to research labs