KABIR MURJANI

AlphaRoute: LLMs as Semantic Optimizers for Multi-Objective Routing[arXiv][Code]

IEEE LAD 2026, Stanford (Oral) · VLSI Routing · ISPD 2025 Benchmarks

• LLM-guided global router where a language model tunes the PathFinder penalties while a knowledge graph keeps every suggestion inside the feasible design space. Repository contains the router, the benchmark harness, and the evaluation scripts.

Drive, Pack, Fly: The Travelling Thief Problem with Drone[arXiv][Code][Data][Weights]

Under review, CAOR · Deep Reinforcement Learning · CORBIT Lab

• Solvers for the TTP-D: exact MILP, simulated annealing and VNS, attention-based construction policies, and a learner-initialised hybrid. Benchmarks and trained weights released alongside on Hugging Face.

Zero-Copy Semantic Contagion[arXiv][Code]

ACM SIGMOD FinDS 2026 (Oral) · Rust · Streaming Architecture

• Full Rust implementation of the zero-copy streaming engine for continuous-time contagion modeling over evolving attention graphs, including the vector similarity search pipeline and benchmarking scripts.

Selective Coverage over Temporal Graphs[Code]

Graph Transformer · PPO/GAE · Submodular Optimization · Collaborator: Nisarg Patel (Google)

• Recast long-context QA as budgeted evidence selection over temporal fact graphs of QMSum transcripts, with succession, contradiction, and relatedness edges, trained under verifiable rewards.
• Improved token-F1 from 25.0% to 27.7% while using 28× fewer input tokens. Proved the coverage objective submodular and NP-hard via Maximum Coverage, giving a 1−1/e greedy guarantee.

Spectral Dominance in Sequential Drug Control[Code]

Zero-Sum Games · Bilevel MILP · Spectral Methods · With Prof. Connor Jerzak (UT Austin)

• Modeled drug scheduling under resistance evolution as a zero-sum game on a within-host compartmental graph, deriving a spectral-radius condition for when sequential treatment dominates myopic dosing.
• Proved single-stage NP-completeness by reduction from 0-1 Knapsack and bounded the myopic–optimal gap across 3 regimes. Validated on 4 calibrated modules (N up to 30 nodes) via a bilevel MILP and a spectral-greedy heuristic.

Aniate: Terminal Command Synthesis via Graph Search[Engram][Website]

Rust · Axum · petgraph · GLiNER + Nomic · PSL Probabilistic Reasoning

• Built Ant, a terminal system mapping natural-language intent to executed commands by searching a graph of candidate command paths, pruned by iterative guardrails and dynamic model routing.
• Designed its context layer, Engram, as a knowledge-graph memory engine over a petgraph-backed entity-relation structure with 256-dimensional semantic embeddings, Rust-accelerated for sub-50 ms retrieval.

Hierarchical RL for Cost–Emission Optimal Dispatch[Code]

Multi-Agent RL · Pareto Frontier · IEEE 118-bus · Advised by Prof. Kuntal Bhattacharjee

• Decomposed 24-hour cost–emission dispatch on IEEE 118-bus (54 generators, 4,242 MW) into a metaheuristic macro layer computing a 15-point trade-off front and a decentralized RL micro layer tracking it in real time.
• Reduced per-zone action space from 5^21 to 7 discrete dispatch levels, making tabular Q-learning tractable while maintaining <0.05% daily fuel-cost variation.

SaNcHaR: Adaptive Topology Routing and Quantized Edge Inference[Code]

1D-CNN · RAG · BLE Simulation · Collaborator: Deep Joshi

• Edge-inference architecture coupling a 1D-CNN with retrieval-augmented generation for localized anomaly detection, alongside a dynamic topology-switching routing layer for degraded networks.
• Achieves an 84.78% packet delivery rate at sub-millisecond latency (0.2778 ms). Repository contains the BLE simulation code and benchmarking scripts reproducing all reported results.

Trained construction policies for the Travelling Thief Problem with Drone — encoder–decoder models that embed the instance once and emit a composite action at each node the truck reaches. Indexed as <variant>/<model>/<family>/n<N>, spanning attention (GAT) and MLP encoders, behaviour-cloned (LISA) initialisers, and both the fixed-endurance (a280) and endurance-conditioned (ttd300) regimes. The benchmark-tuned policies warm-start from sampled/n20, as training from scratch does not converge within budget at larger sizes.