Kabir Murjani

Code

Working Projects

ChessNano: Sub-50M Parameter Autoregressive Transformer

PyTorch · GQA · INT8 Quantization · OpenAI Parameter Golf

Engineered for OpenAI's global optimization challenge, designing and executing a quantized transformer architecture to maximize compression performance.
Selected as a grant recipient by OpenAI, securing the 8xH100 infrastructure allocation to deploy, train, and evaluate the architecture.
Adhered to L(N) constraints, ensuring the model operated strictly within an absolute 16MB file threshold and a 10-minute training wall-clock on an 8xH100 cluster.
Architected a GQA backbone with TurboQuant quantization-aware training (Hadamard PolarQuant + 1-bit QJL correction) and Tree-of-Thought constrained decoding, predicting moves from Standard Algebraic Notation (SAN) without explicit board representation.

Projects

Cognitive Adjudication Layer (CAL)

Markov Random Fields · ADMM · Graph Theory · Collaborator: Nisarg Patel (Google)

Architected a neuro-symbolic, multi-hop reasoning engine executing over a 15M-edge financial knowledge graph derived from SEC filings, utilizing semantic anchor discovery and RustworkX for bounded-depth bi-directional pathfinding.
Mathematically formulated and solved a Hinge-Loss Markov Random Field (HL-MRF) via ADMM (Alternating Direction Method of Multipliers), utilizing Lukasiewicz T-norms for structural adjudication and inference-time calibration.

Publications & Implementations

SaNcHaR: Adaptive Topology Routing and Quantized Edge Inference

IEEE ICC, 2026 · 1D-CNN · RAG · BLE Simulation · Collaborator: Deep Joshi

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

Zero-Copy Semantic Contagion

ACM SIGMOD FinDS, 2026 · Rust · WebAssembly

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

Weight-Space Teleportation: Discovering LLM Reasoning via Bilevel Evolution

TMLR (Under Review) · Collaborator: Parth Vyas

Bilevel framework coupling SVD-stratified mutations with gradient refinement, reaching 77.6% on GSM8K. Repository contains training scripts, mutation operators, and evaluation harness.

Judges Hallucinate, Embeddings Don't (JADE)

TMLR (Under Review) · Collaborator: Parth Vyas

Judge-free alignment framework using deterministic embedding regression to improve human preference by 44%. Repository contains the JADE pipeline, preference dataset construction scripts, and evaluation code.

Datasets

CRAFT-5

Constrained Reasoning and Adaptive Feedback Training

A high-quality RLAIF dataset designed for training and evaluating language models on constrained problem-solving tasks with multi-dimensional quality assessment. Curated for academic research. Contains 2,384 examples of instruction-following tasks with comprehensive constraint handling, detailed reasoning traces, and multi-dimensional quality ratings (1–5 scale) suitable for preference learning and RLHF applications.

View CRAFT-5 on Hugging Face

All repositories are installable as a single Python package with pip install kabir and importable by name, e.g. from kabir import chessnano.