I build reliable software and explain it clearly.

Implemented and compared three reinforcement learning algorithms: Value Iteration, Q-Learning, and SARSA, to understand convergence, stability, and policy quality under stochastic dynamics and crash penalties.

• Modeled a high dimensional state space combining position and velocity, with stochastic acceleration failures and two crash semantics (restart vs nearest cell).
• Implemented Value Iteration for exact planning and Q-Learning/SARSA for model free learning, comparing convergence behavior, policy quality, and computational cost.
• Evaluated learned policies across multiple track geometries (2-track, W-track, U-track), highlighting tradeoffs between optimality, stability, and runtime.

Built a flexible Sudoku solving framework implementing multiple search and constraint satisfaction algorithms to compare pruning strength, search behavior, and decision level efficiency (not wall clock time).

• Implements five solving strategies (backtracking, forward checking, AC-3, simulated annealing, genetic algorithms) behind a shared constraint checking interface.
• Uses constraint propagation and pruning to dramatically reduce search space compared to naive backtracking approaches.
• Compares algorithm behavior using decision level metrics to enable fair, interpretable evaluation across deterministic and stochastic methods.

Built a First-Order Logic reasoning system that classifies a query cell as SAFE, UNSAFE, or RISKY by turning percepts into constraints and proving conclusions via contradiction.

• Translates path percepts (breeze/stench) and domain rules into CNF clauses, builds a knowledge base, then attempts proofs via contradiction for SAFE/UNSAFE.
• Implements core inference machinery (unification + resolution) independently of the domain rule set, keeping Wumpus specific rules separate from the reasoning engine.
• Evaluated on 12 caves with an overall accuracy of 83% (10/12); observed inference cost increasing sharply with grid size and percept ambiguity.

Implemented and evaluated Variable Elimination (exact) and Gibbs Sampling (approximate) to quantify the accuracy/runtime tradeoff as Bayesian networks scale in size and evidence complexity.

• Built VE and Gibbs Sampling from scratch, including factor operations, elimination ordering, sampling schedules, and normalization, without external inference libraries.
• Evaluated accuracy, runtime, and scalability on three networks (Child, Insurance, Win95pts), showing VE achieves near perfect accuracy while GS scales far better computationally.
• Demonstrated the real world tradeoff between exact inference and stochastic approximation, confirming GS as the practical choice for large, complex networks.