Adarsh Kumar

An agentic payment system that lets an LLM agent pay for API access over the x402 protocol against an on-chain spending budget it can never read, alter, or override.

• Deployed the encrypted on-chain budget (Inco Lightning) live on Base Sepolia behind a 327-test CI suite, capping a compromised agent's worst-case loss by policy.
• Specified TEE-derived signing keys (Oasis ROFL) as the custody model — designed, not yet deployed.
• Enforced a CI-checked import boundary blocking the untrusted orchestrator from reaching the signer or the vendor's API credential.
• Committed each spend's debit before the approval decision was knowable, closing the obvious race.

A top-1 routed mixture-of-experts language system built from a 135M-parameter base, with three domain experts and a learned router.

• LoRA fine-tuned three domain experts against frozen attention weights — about 1M trainable parameters per adapter, 0.74% of the base.
• Trained a router MLP over 384-dimensional sentence embeddings on ~7,000 generated queries to dispatch each query to a single expert.
• Evaluated with router accuracy, a confusion matrix, and per-expert perplexity rather than a single aggregate score.

A C++ library pricing European and American options under four models, with a validation suite that makes the models check each other.

• Implemented Black-Scholes (closed-form), Binomial Tree (backward induction), Monte Carlo (100K-path simulation), and Heston (stochastic volatility).
• Wrote a validation suite checking put-call parity, American ≥ European price under early exercise, and intrinsic-vs-time-value decomposition.

A competitive AI for the game Chain Reaction, tuned to search as deep as it can inside a hard move budget.

• Minimax with alpha-beta pruning, iterative deepening, and a transposition table.
• Tuned to reach maximum depth inside a 0.9-second per-move budget.