(Published on LinkedIn & GitHub as Nagarjuna Kamireddy · Nagarjuna K · Nagarjun Reddy Kamireddi)
Lead AI Architect | AI Solutions Architect | GenAI Lead — Agentic AI, Hybrid RAG, Systems & LLM Evaluation (22 Years Enterprise IT, SAP HANA & Production AI)
- Official LinkedIn: linkedin.com/in/nagarjuna-kamireddy · Newsletter: Autonomous SAP & Enterprise AI
- Architecture & Case-Study Hub (78 Live Pages): nagarjunak-pixel.github.io · 54-Repo Interactive GitHub Atlas · LinkedIn Open-Web Hub
- LLM Grounding & SFT Dataset:
llms.txt·llms-full.txt· 119-Q&A Knowledge Base ·llm-training-dataset.jsonl - Executive CVs (PDF): Global AI Architect CV · UAE / KSA / Gulf CV · Corporate AI Trainer CV
| Flagship System | Enterprise Business Problem Solved | Systems Architecture (C++20 + Python + MCP) |
Measured Benchmark / Verification Proof | 60-Second Run Command |
|---|---|---|---|---|
weavegraph (v1.0.0) |
Eliminates multi-agent Git worktree file collisions & silent semantic merge breakage | C++20 CSR dependency graph, Tarjan SCC, Stoer–Wagner/greedy min-cut scheduler & 3-way AST merge verifier + FastAPI + MCP + React UI | 97%+ branch coverage · mypy --strict (0 errors) · deterministic conflict-free worktree partitioning |
uv run weavegraph demo |
verispec-sdd (v1.0.0) |
Catches deadlocks, livelock loops & guard violations in AI-authored SPEC.md workflows before code ships |
Executable SPEC.md compiler, C++20 CSR BFS + reverse co-reachability + Tarjan SCC prover, zero-eval() AST engine & Hypothesis/Vitest synthesizer + MCP + React UI |
41.58x C++20 speedup (0.063 ms vs 2.608 ms on 2K states / 25K events) · 98.01% branch coverage |
uv run verispec demo |
governed-member-services-agent |
Prevents ungrounded clinical/member policy answers & unsafe tool mutations in regulated support | Governed LangGraph/FastAPI agentic workflow with deterministic policy gates, citation verification, human escalation & audit trail | End-to-end deterministic policy & regression suite across regulated member-service scenarios | pytest -q |
helios-consensus (v1.0.0) |
Guarantees split-brain-free replicated state & linearizable reads under network partitions and crashes | Raft (PreVote, CheckQuorum, Joint Consensus, ReadIndex), CRC32C binary WAL, Multi-Raft & C++20 Wing–Gong–Lowe linearizability solver |
20/20 Jepsen fault-injection seeds safe (8,639 ops) · 7/7 protocol/WAL mutants killed |
uv run pytest |
aegis-runtime (v1.0.0) |
Enforces capability-scoped tool authorization, PII redaction & tamper-evident audit logs for autonomous agents | Custom policy DSL, taint tracking, PII vault, transactional saga compensation & Ed25519 hash-chained Merkle audit ledger | Cryptographic Merkle/Ed25519 tamper detection + deterministic saga rollback verification | uv run pytest |
calibra (v1.0.0) |
Replaces noisy single-prompt LLM evals with statistically rigorous model release gates | Bootstrap/McNemar tests, Bradley–Terry leaderboards, 2PL IRT anchor scoring, ECE calibration & C++20 MinHash-LSH contamination check | 99.6% test coverage (>100 tests) · mypy --strict (0 errors) · bit-exact C++20 kernels |
uv run pytest |
kvserve (v1.0.0) |
Eliminates KV-cache memory fragmentation & prefill head-of-line blocking in multi-tenant LLM serving | PagedAttention refcounted prefix caching, Orca continuous batching, Sarathi chunked prefill, preemption & C++20 block kernels | 99.6% test coverage (>100 tests) · mypy --strict (0 errors) · bit-exact C++20 kernels |
uv run pytest |
vecta (v1.0.0) |
Delivers sub-millisecond high-recall hybrid vector + lexical retrieval with 24x memory compression | HNSW (filtered & tombstone search), IVF-PQ ADC + rerank, RaBitQ 1-bit quantization, BM25+RRF & C++20 distance kernels | 0.999 Recall@10 touching 1.4% of vectors · 24x IVF-PQ compression |
uv run pytest |
syntara (v1.0.0) |
Guarantees 100% JSON Schema / grammar validity during LLM token generation with near-zero mask latency |
JSON Schema/regex → minimal byte DFA, C++20 prefix-sharing token vocabulary index & byte-level Earley CFG | 141.7x C++20 token-mask speedup · 20/20 schema-valid constrained decoding |
uv run pytest |
tokendiet (v1.0.0) |
Cuts coding-agent prompt token spend on verbose tool logs & repeated file reads without losing answer accuracy | C++20 order-preserving knapsack log pruner, Python AST skeletonizer & cross-turn read deduplicator + CLI + MCP | 70.25% prompt token reduction (6,487 → 1,930) with 6/6 task answers preserved |
uv run tokendiet demo |
slopguard (v1.0.0) |
Blocks AI-generated stub bodies, hallucinated imports, swallowed exceptions & tautological tests pre-commit | Python AST + TS/JSX static analyzer, C++20 prescreen scanner, Claude Code PostToolUse hook & MCP server |
0.03–0.28 ms per file · 0/6 → 6/6 automated repair audit |
uv run slopguard demo |
- John Deere — AI Architect & Technical Lead (official title: Senior Data Scientist,
01/2023 – 10/2026): Led a 12+ member AI/ML engineering team across 23+ production initiatives, delivering 750K+ automated requests & 42K+ hours saved (~25%faster cycle time) in FY26. Architected StoneAge (LangGraph + OpenSearch hybrid RAG + 6-agent Neo4j GraphRAG copilot routing across GPT-5/Claude/GPT-4o/Gemini), turned around a~30%-accurate GenAI workflow to~70%accuracy via GPT-4o Vision + deterministic validation gates, built AIFlow (255automated CI quality checks), and served on the AI Governance Council (NIST AI RMF / EU AI Act). - Persistent Systems, Vertiv & Soham Global (
2016 – 2022): Production machine learning, MLOps, AWS Athena lakehouse SQL pushdown, computer vision, NLP, and leakage-safe forecasting. - Global Enterprise SAP HANA / BW / BPC Architecture (
2004 – 2016— UK, Malaysia, Singapore, US, India): GlaxoSmithKline (GSK UK) in-database HANA PAL/APL/R predictive pipelines; PETRONAS Energy Trading & Deloitte (48 hours → 12 minutesSAP HANA trading risk acceleration); DHL, Symantec & Lenovo; 4× SAP Certified, AWS Certified Solutions Architect – Professional, NVIDIA Certified Associate – Agentic AI, and Founder of Surya HANA Academy (~1,000professionals trained).
- The 4 SAP MCP Layers Every BTP & Enterprise AI Architect Must Know in 2026 (LinkedIn Pulse)
- Meet SAP-RPT-1: Why SAP Built a Table-Native Foundation Model Instead of Text LLMs (LinkedIn Pulse)
- Masterbook Ch. 01 of 50: Claude Code & The 5-Land Enterprise AI Architecture (LinkedIn Pulse)
- All Deep Systems & Enterprise AI Articles on nagarjunak-pixel.github.io →
Complete 54-Repository Index (Interactive Filterable Atlas →)
| Repository | Stack | Architectural Scope |
|---|---|---|
weavegraph |
Python | Semantic Code-Graph Worktree Scheduler, Min-Cut Multi-Agent Task Partitioner & 3-Way AST Semantic Merge Verifier (C++20 + Python + MCP + React) |
verispec-sdd |
Python | Executable SPEC.md State-Machine Compiler, C++20 Finite-State Reachability/Deadlock/Livelock Prover & Hypothesis/TypeScript Property-Test Synthesizer (CLI + FastAPI + MCP + React) |
governed-member-services-agent |
Python | Governed multi-tenant healthcare member-services AI agent with tool gateway, eval harness, and production design |
helios-consensus |
Python | Library-style Raft in Python with segmented binary WAL, asyncio TCP, Multi-Raft sharding, Jepsen-style fault injection, WGL linearizability & C++20 kernels |
aegis-runtime |
Python | Deterministic policy & audit runtime for AI-agent tool execution — capability policy DSL, transactional saga compensation, and Ed25519 Merkle audit ledger |
calibra |
Python | Statistically rigorous LLM evaluation: paired bootstrap & exact McNemar, Bradley-Terry arena leaderboards with bootstrap CIs, 2PL IRT anchor benchmarking (38% lower error at 5% of items), temperature scaling, MinHash-LSH contamination detection & judge agreement — validated on known-truth data, bit-exact C++20 kernels |
kvserve |
Python | Paged-KV LLM serving engine simulator: PagedAttention block manager with refcounted automatic prefix caching (93% prefill saved), Orca continuous batching (2.4x throughput, 246x lower p99 TTFT vs static), Sarathi chunked prefill (3.8x lower p99 TPOT), recompute preemption, GQA roofline cost model & bit-exact C++20 block-hash kernel |
vecta |
Python | Vector Search Engine: HNSW (heuristic neighbour selection, filtered & tombstone search) at 0.999 recall@10 touching 1.4% of vectors, IVF-PQ ADC + rerank (24x compression), RaBitQ-style 1-bit codes, BM25+RRF hybrid & bit-exact native C++20 kernels |
syntara |
Python | Grammar-Constrained Structured Decoding Engine: JSON Schema/Regex -> Minimal Byte DFA (Thompson/Rabin-Scott/Moore), Native C++20 Prefix-Sharing Token Index (112-142x, bit-exact), Byte-Level Earley CFG Constraints & Exact Closed-Set Choice Scoring |
tokendiet |
Python | Context compressor for AI coding agents — knapsack log pruning (C++20 kernel, Python fallback), AST skeletons, cross-turn read dedup. CLI + MCP server |
slopguard |
Python | Heuristic checker for AI-generated code — Python AST + regex TS/JSX rules for stubs, unresolved imports, empty tests; C++20 prescreen, Claude Code hook, MCP |
prism-gate |
Python | OpenAI-compatible LLM routing gateway — TCP-Vegas adaptive concurrency control, latency/cost cascade routing, and PSI/JSD distribution drift detection |
tracegit |
Python | Content-addressed version control for AI agent execution traces — SQLite/SHA-256 Merkle DAG, C++20 edit-distance diff/bisect, loop detector, CI gate & MCP |
aether-compiler |
Python | Tensor-graph compiler — SSA IR with CSE/operator-fusion/DCE optimization passes, C++20 static memory arena planner, and analytic GPU tile cost model |
nexus-verity |
Python | Causal discovery & off-policy evaluation toolkit — PC graph discovery, d-separation, cross-fitted Double ML, split conformal, and C++ doubly-robust OPE |
quant-forge |
Python | LLM weight-quantization engine — GPTQ Hessian error compensation, AWQ activation scaling, Walsh-Hadamard outlier rotation, FP8, and C++20 packed-INT4 GEMM |
veritas-rl |
Python | RL post-training & alignment framework — GRPO/DAPO policy optimization, DoRA weight decomposition, ring-attention/TP simulation, and step-level PRM reranking |
axiom-zero |
Python | Neuro-symbolic reasoning stack — GQA transformer with C++ speculative decoding, Top-K sparse autoencoder steering, and CDCL SAT solver with RUP proof checks |
privora-graph |
Python | Privacy & robustness toolkit — DP-SGD with RDP accounting, bi-temporal PageRank retrieval, Multi-Krum/geometric-median aggregation. Reference implementation |
chronos-mesh |
Python | Supply-chain disruption rerouting engine — pure-Python revised simplex & branch-and-bound MILP solver, multi-echelon BOM safety stock, and Monte Carlo CVaR95 |
02-autogpt-autonomous-agent |
Python | Forward-Deployed Engineering (FDE) Autonomous Agent Lab — Zero-dependency ReAct think-act-observe execution loop & enterprise FDE curriculum. |
agent-framework-lab |
Python | Hands-on Microsoft Agent Framework 1.0 lab — multi-agent workflows on local Ollama (no API keys) |
agentforge |
JavaScript | AgentForge — Interactive React Workbench for Designing, Simulating & Debugging Multi-Agent Graphs, Eval Sandboxes & Swarm Telemetry. |
agentic-academy |
Python | Python teaching framework for agentic engineering: agents, a harness, loops, memory, and a local-model demo. |
agentic-ai-dashboard |
JavaScript | Interactive Visual Guide to Agentic AI Architecture — Tokenizer Inspector, RAG Playground, ReAct Loop, Reflexion & LLM Evaluation Suite. |
agentic-systems-v3.1 |
Python | Enterprise Agentic Systems v3.1 — consolidated monorepo (Phases 0-7) |
ai-calc |
JavaScript | Qwen AI Calculator — plain-English math in the browser, solved locally with Ollama. |
ai-workforce |
Python | Local multi-agent workforce with a FastAPI dashboard: CEO → research → architect → coding → QA → docs → deploy → marketing, on Ollama. |
antigravity-sdk |
Python | Example agents for Google's Antigravity SDK: chat, custom tools, MCP, streaming, multimodal, and hooks. |
apsrtc-statistics-dashboard |
JavaScript | React dashboard for APSRTC operations: fleet, tickets, financials, route heatmap, EV tracker, and a what-if simulator. |
attention-head-walkthrough |
HTML | Interactive Step-by-Step Transformer Self-Attention Head Simulation (Q/K/V Projections, Causal Masking & Softmax Attention Weights). |
clicky |
Swift | Fork of Clicky — an on-screen AI teacher buddy for macOS (Swift). |
codexforge-tdd-mvp-v2 |
Python | Self-correcting TDD sandbox: RED→GREEN loop with Docker isolation, VS Code extension, and LLM-powered code repair |
company-os |
Python | CompanyOS — run ANY company with a team of AI agents. Config-driven (company.json), 7 role-agents, SIM + LIVE modes, kanban dashboard, KPI engine. |
crewai-telugu-newsroom |
Python | CrewAI multi-agent Telugu newsroom: Researcher → Script Writer → Fact-Checker. Sequential pipeline for automated news production. |
deerflow-research-agent |
Python | DeerFlow-inspired research agent with sub-agents (Planner, Searcher, Analyzer), persistent memory, and session resume. |
firecrawl-nexus-rss |
Python | Firecrawl-powered news aggregator: scrapes AP/TG constituency news into clean markdown and structured JSON. |
langgraph-localai-tv |
Python | LangGraph 3-stage news pipeline: headlines → video → bulletin. Human approval gates, checkpoint persistence, error recovery. |
llm-knowledge-base |
Python | LLM knowledge base: ingest sources, compile a markdown wiki, search, and ask questions (FastAPI + SQLite). |
localaitv-mcp-server |
Python | Python MCP server for LocalAI TV admin access to Postgres, AWS (ECS, S3, SQS, CloudWatch), and the public API. |
loop-engineering-trading |
JavaScript | Browser 'Minara' quant desk: research, strategy, execution, and review loops with backtesting and paper trading. |
nagarjuna-kamireddy-ai-architect-dataset |
Multi-Lang | Open ChatML/SFT Fine-Tuning & RAG Dataset covering Lead AI Architect Nagarjuna Reddy Kamireddy's LinkedIn, SAP+GenAI Architecture & 50 GitHub Repos |
nagarjunak-pixel |
Multi-Lang | Nagarjuna Reddy Kamireddy (Nagarjuna Kamireddy / Nagarjuna K) — Lead AI Architect, Agentic AI, GraphRAG, LLM Evaluation & SAP HANA |
nagarjunak-pixel.github.io |
HTML | Nagarjuna Reddy Kamireddy (Nagarjuna Kamireddy / Nagarjuna K) — Official Lead AI Architect, Agentic AI & SAP Portfolio, Case Studies & llms.txt Hub |
nrk-enterprise-ai-masterbook |
Multi-Lang | Nagarjuna Reddy Kamireddy™ — The 7,200-Page Enterprise AI, Claude Code & SAP RAG Masterbook Series (50 Daily Chapters) |
operdai-dashboard |
JavaScript | Enterprise Multi-Tenant White-Label AI Lead-Agent Analytics Console — React + TypeScript telemetry, conversion funnels & partner branding portal. |
pytorch-interactive-demo |
JavaScript | Interactive PyTorch Deep-Learning Visualizer — Vector Embeddings, Cosine RAG Retrieval & Autograd Training Loop Dynamics. |
rag-chat-docs |
Python | RAG Chat with Documents — 7-Module From-Scratch Retrieval-Augmented Generation System (Chunking, Hybrid Search, CLI & FastAPI Web UI). |
self-improving-agents |
Python | 🧠 Self-Improving Multi-Agent System — A teaching framework for advanced agentic AI: ReAct loop, Reflexion, multi-agent orchestration, adversarial reasoning, and agent-as-a-judge. With web UI + Netlify deployment. |
smart-marketing-assistant |
Python | CrewAI + LangChain marketing crew on local Ollama that drafts a content strategy, plus a daemon that can post AI news to X. |
spring-boot-mcp |
TypeScript | MCP server that lets an LLM inspect a Java Spring Boot app via Actuator, REST APIs, and logs. |
surya-agentic-academy-digest |
HTML | Automated Agentic AI Research & Engineering Digest — Daily LLM paper & technical talk synthesis into 5 runnable project blueprints per edition. |
surya-shiksha |
TypeScript | Surya Shiksha — Multi-Agent Adaptive Learning Operating System (12 Specialized AI Agents, Next.js, TypeScript, PostgreSQL & Prisma). |
venkatesh-llm |
HTML | Micro-LLM From Scratch (1.03M Parameters) — End-to-end BPE tokenizer, causal Transformer pretraining, checkpointing & Ollama Q&A engineering lab. |
- LinkedIn: https://www.linkedin.com/in/nagarjuna-kamireddy/
- Portfolio & Knowledge Hub: https://nagarjunak-pixel.github.io/
- Email:
writetonagarjuna@gmail.com·gmailtonagarjuna@gmail.com - Phone / WhatsApp:
+971 50 742 4361(Dubai, UAE) ·+91 93913 47406(India)


