github.com/nagarjunak-pixel · 38 Public Repositories · 2,249+ Annual Contributions

Open-Source Frontier & Enterprise AI Architecture Suite

Runnable open-source systems built by Nagarjuna Reddy Kamireddy spanning post-training RLVR, speculative transformer runtimes, zero-trust ERP agent control planes, multi-tenant LLM gateways, multi-agent products, and MCP integrations.

Sovereign Privacy, Bi-Temporal GraphRAG & Byzantine FL

privora-graph

Exact Binomial Rényi Differential-Privacy (DP-SGD) accountant (ε = 4.14, -90.3% MIA leakage), SQL:2011 bi-temporal Knowledge Graph with native C++20 CSR Personalized PageRank (0/6 retroactive compliance leaks), and Multi-Krum / Weiszfeld Byzantine-resilient federated learning.

DP-SGD / RDPBi-Temporal GraphRAGC++20 CSRByzantine FL
Frontier RLVR & Parallelism

veritas-rl

Critic-free GRPO/DAPO Verifiable-Reward RL engine with native C++20 advantage kernels, DoRA weight decomposition (ICML '24 Oral), causal Online-Softmax Ring-Attention + Megatron 2D tensor parallelism, and live Step-Level Process Reward Model (PRM) test-time compute.

PyTorchC++20GRPO / DAPODoRA2D TP
Verified Transformer Runtime & Interpretability

axiom-zero

From-scratch RoPE + GQA + SwiGLU Transformer with CoW Paged-KV cache, native C++20 lossless speculative decoding kernel, Anthropic-style Top-K Sparse Autoencoder (SAE) causal circuit steering, and 1-UIP CDCL RUP formal proof verification.

Speculative DecodingPaged KVTop-K SAECDCL / SMT
Sub-Byte LLM Compression & Roofline Engine

quant-forge

Native C++20 bit-packed 2-weights/byte INT4 register-fused GEMM (6.40x compression), Second-Order Cholesky GPTQ (25.49x lower MSE vs RTN), Salient-Channel AWQ, SpinQuant Walsh-Hadamard outlier rotation, OCP FP8 E4M3/E5M2, and H100/M3 Max Roofline autotuner.

C++20 INT4Cholesky GPTQAWQSpinQuant FWHTFP8
Zero-Trust Enterprise Agent Control Plane

aegis-runtime

Deterministic zero-trust control plane for enterprise AI agents featuring a formal AST policy compiler, Biba/Bell-LaPadula taint lattices, AES-256-GCM PII vaulting, transactional SAP/ERP compensating rollbacks, and RFC-6962 Merkle + Ed25519 audit ledgers.

Zero-TrustERP SagasMerkle AuditAgent Safety
Multi-Tenant LLM Inference Gateway

prism-gate

High-throughput OpenAI-compatible multi-tenant LLM gateway implementing TCP-Vegas adaptive concurrency limiting, speculative cascade model routing, inline atomic claim-faithfulness judging, and PSI + Jensen-Shannon statistical drift monitoring.

LLM GatewayFinOps RoutingPSI / JSD DriftPython
Neuro-Symbolic Operations Research

chronos-mesh

Multi-echelon supply-chain digital twin combining exact Branch-and-Bound Mixed-Integer Linear Programming (MILP) rerouting across Cost/SLA/Scope-3 Carbon Pareto frontiers with a Monte Carlo CVaR95 tail-risk engine.

Neuro-SymbolicMILP SolverSupply ChainCVaR95
50-Chapter Enterprise Architecture Publication

Enterprise AI, Claude Code & SAP RAG Masterbook

Comprehensive 50-chapter enterprise reference series covering production agentic architectures, SAP HANA / BTP RAG integration, MCP tool servers, evaluation harnesses, and enterprise AI governance.

Enterprise AISAP RAGMCP50 Chapters
Autonomous Organization Engine

company-os

Config-driven autonomous company runtime where 7 executive AI agents plan, assign, and track cross-functional work from a single company.json specification — complete with Kanban board, live streaming logs, SIM/LIVE modes, and KPI engine.

Multi-AgentPythonKanbanKPI Engine
Agentic EdTech Platform

surya-shiksha

Agentic learning operating system powered by 12 specialized GPT-4o agents handling adaptive curriculum generation, Socratic tutoring, mastery tracking, and formative assessment.

Next.jsTypeScriptPostgreSQLPrisma12 Agents
Self-Maintaining LLM Wiki & Search

llm-knowledge-base

Ingests raw heterogeneous sources and compiles an LLM-maintained Markdown knowledge wiki backed by SQLite FTS5 full-text search, FastAPI endpoints, and grounded Q&A synthesis.

FastAPISQLite FTS5ClaudeRAG Wiki
Stateful Media & Research Pipelines

langgraph-localai-tv & crewai-telugu-newsroom

Production-style stateful LangGraph 3-stage media pipeline (headlines → video → bulletin) with checkpoint persistence, interrupt/resume recovery, and mandatory human approval gates, alongside a CrewAI Researcher → Writer → Fact-Checker pipeline.

LangGraphCrewAIHITL ApprovalCheckpointing
Agentic Engineering Workbenches

agentforge, self-improving-agents & agentic-ai-dashboard

Interactive React & Python teaching frameworks covering visual multi-agent graph construction, ReAct/Reflexion loops, adversarial reasoning, agent-as-a-judge evaluation sandboxes, and RAG playgrounds.

ReactReflexionEval SandboxAgent-as-Judge
Deep Learning From First Principles

venkatesh-llm & attention-head-walkthrough

End-to-end 1.03M-parameter educational language model (dataset → BPE tokenizer → PyTorch transformer training → Ollama serving) paired with an interactive step-by-step transformer attention-head simulator.

PyTorchTransformerTokenizerOllama