Skip to content
View Akgithub2028's full-sized avatar
🤖
🤖
44 followers · 49 following

Highlights

  • Pro

Block or report Akgithub2028

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Akgithub2028/README.md
Typing SVG

AI Systems & Agentic Infrastructure Engineer · Agentic AI Research · Distributed Cloud Systems
Indian Institute of Technology (IIT) Kharagpur

Followers LinkedIn Portfolio ArXiv Preprint

github contribution snake animation

(Snake updates automatically from live contribution activity)

🌌 About Me

I am an AI Systems & Agentic Infrastructure Engineer and researcher at the Indian Institute of Technology (IIT) Kharagpur. My work focuses on building mathematically grounded,agentic runtimes, speculative execution engines, protocol security defense platforms, and enterprise distributed cloud backbones.

I bridge the gap between academic AI research and high-throughput production infrastructure:

  • 🔬 Agentic AI Research & Invariance Theory: Formulated the Tool-Trust Invariance benchmark and empirical preprint, establishing formal metrics ($\text{CTSS}$, $\text{SFR}$) to diagnose non-invariant tool selection in foundation models, backed by zero-cost deterministic canonicalization middleware.
  • ⚡ Speculative Agent Execution & Runtimes: Engineered SpecTool, a zero-dependency speculative execution runtime with Python AST verification, effect-system safety, and transactional rollback that hides agent invocation latency across chained tools.
  • 🏢 Enterprise Decision Intelligence: Architected NexusAgent, a production-grade multi-agent platform combining clause-aware hybrid RAG (Milvus dense + BM25 sparse + BGE cross-encoder) and MCP NL2SQL with AST-level SQLGlot guardrails.
  • 🕸️ A2A Dynamic Orchestration: Developed LOOM, a parallel multi-agent DAG compiler over Google GenAI streaming, achieving $2.06\times$ topological speedup and failure-localized replanning across Agent-to-Agent (A2A) microservices.
  • 🛡️ MCP Security & Sub-Millisecond Defense: Built MCP-Sentinel, an enterprise red-teaming and runtime defense substrate for the Model Context Protocol, defending against prompt injection, unauthorized tool invocation, and privilege escalation.
  • ☁️ Cloud Infrastructure & Production Engineering: Orchestrated microservices across Google Cloud Platform (GCP - Vertex AI, GKE, Cloud Run), Amazon Web Services (AWS - Bedrock, EKS), and Kubernetes (K8s) with OpenTelemetry observability, Helm, Docker, and Redis 7 / PostgreSQL 16 backbones.

🛠 Enhanced Tech Stack & Systems Architecture

🤖 Agentic AI & Systems Engineering






Autonomous multi-agent orchestration, typed tool schema contracts, speculative tool execution, A2A communication protocols, and hybrid RAG.

☁️ Cloud & Container Infrastructure






Production microservice autoscaling, multi-cloud LLM gateways, declarative IaC with Terraform, connection pooling, and distributed consensus.

🛡️ AI Reliability, Post-Training & Security






Semantic LLM red-teaming, AST query gates, transaction reconciliation, fail-stop safety policies, and task-clustered bootstrap uncertainty.

⚙️ Core Systems & Software Engineering





High-concurrency async services, strict Pydantic v2 schema validation, systems-level lock-free memory management, and reproducible CI/CD pipelines.

🚀 Flagship AI Systems, Runtimes & Research

🔬 Tool-Trust Invariance Lab


Formal empirical benchmark and publication paper evaluating whether frontier LLMs (Llama-3.2-11b, GPT-OSS-20B) select tools invariantly under semantics-preserving catalog transformations.

  • Formulation: Formulated Conditional Tool Selection Stability ($\text{CTSS} = 96.3%$) and Selection Flip Rate ($\text{SFR} = 24.3%$).
  • Intervention: Zero-cost deterministic catalog canonicalizer recovering +3.0% stability without answer leakage or token inflation.
  • Rigorous Statistics: 1,000-resample task-clustered bootstrap uncertainty estimation; complete two-column arXiv LaTeX package with 7 publication figures.

⚡ SpecTool


High-performance speculative tool-execution runtime for autonomous AI agents with conservative dependency and effect verification.

  • Latency Hiding: Employs conservative dependency graph analysis to speculatively dispatch non-blocking tool calls ahead of LLM token completion, achieving a 1.25× end-to-end latency reduction.
  • Effect-System Safety: Validates speculative calls via Python AST parsing to guarantee pure side-effect-free execution.
  • Transactional Reconciliation: Implements canonical invocation identities promoting exact-match speculative results safely with zero rollback leaks.
  • Rigorous Testing: Formally verified through Hypothesis property-based testing and deterministic edge-case suites.

🏢 NexusAgent — Enterprise Decision Intelligence


Enterprise multi-agent decision intelligence platform powered by LangGraph, clause-aware hybrid RAG, and MCP NL2SQL with AST-level safety guardrails.

  • Hybrid RAG Pipeline: Fuses Milvus dense vectors, BM25 sparse lexical retrieval, Reciprocal Rank Fusion (RRF), and BGE cross-encoder reranking, attaining 96.25% Hit@1 accuracy.
  • AST SQL Guardrails: Enforces AST-level query validation via SQLGlot with tenant-scoped, read-only gating over MySQL databases.
  • Production Architecture: High-concurrency FastAPI orchestration layer with 1,802 automated unit/integration tests and six strict CI quality gates.

🕸️ LOOM


Dynamic Agent-to-Agent (A2A) task graph generation, subtask independence verification, and parallel multi-agent orchestration runtime.

  • LLM-to-DAG Compilation: Compiles natural language intent into dependency-aware DAGs with strict Pydantic v2 structured schemas over Google GenAI streaming.
  • Topological Scheduling: Achieved a 2.06× parallel speedup via asyncio semaphore-bounded scheduling across distributed A2A worker nodes.
  • Localized Fault Tolerance: Implements failure-localized replanning that isolates faulted nodes and dynamically replans without corrupting completed subgraph states across Docker containers.

🛡️ MCP-Sentinel: Security Auditing & Runtime Defense


Enterprise security auditing, semantic LLM inspection, and sub-millisecond runtime defense platform designed specifically for the Model Context Protocol (MCP).

  • Sub-Millisecond Interception: High-throughput gateway interceptor inspecting tool call payloads and responses with sub-millisecond overhead.
  • Adversarial Red-Teaming: Multi-layer defense combating prompt injection, privilege escalation, unauthorized tool execution, and toxic memory poisoning.
  • Automated Auditing: Static capability analysis and runtime telemetry yielding standardized SARIF vulnerability reports for DevSecOps pipelines.
📦 Additional AI Infrastructure, Inference & Multi-Agent Repositories (Click to Expand / Collapse)
Repository Focus Area Key Architecture & Capabilities Primary Language
Agent-Reliability-and-Evaluation-Lab MCP Runtime Reliability Production reliability substrate for MCP runtimes: circuit breakers, protocol negotiation, adaptive streaming, connection pooling, and 109 automated tests. Python 3.12
Agent-Harness Deterministic Agent Eval Deterministic evaluation sandbox benchmarking multi-turn agents on trajectory adherence, tool precision, and safety policies. Python 3.12
nano-VLLM LLM Serving from Scratch Minimal high-performance inference engine built from scratch implementing PagedAttention, continuous batching, and KV-cache blocks. Python / CUDA
nano-llm-posttraining Post-Training & Alignment Minimal alignment on 8GB GPU: KL constraints, SFT, DPO, and DeepSeek-R1-style GRPO reasoning distillation. Python / PyTorch
Autonomous-AI-Co-Scientist Multi-Agent Cognitive Platform Autonomous research system implementing and extending Google's AI Co-Scientist with persistent vector memory and containerized Docker pipelines. Python
AutoResearch Self-Reflective Agents Hierarchical supervisor-worker agent topology for autonomous scientific hypothesis generation and peer review. Python
VoRTeX AI Workflow Execution Plane Event-sourced execution engine with PostgreSQL checkpointing, multi-provider LLM gateway, and Redis rate limiting (182.8 req/s). Python / FastAPI
Distributed-In-Memory-Key-Value-Store Distributed Systems & Storage High-throughput distributed in-memory cache and key-value database built in modern C++ with consensus replication and sub-millisecond retrieval. Modern C++
📊 Quantitative Systems & High-Frequency Trading Projects (Click to Expand)
Repository Focus Area Key Architecture & Metrics Language
Real-Time-Market-Making-Execution-Engine Low-Latency C++ HFT Production market-making engine: lock-free SPSC queues (1.00µs p50), Avellaneda-Stoikov inventory control, and L2/L3 order book reconstruction. C++20
Multi-Alpha-Engineering-Convex-Optimization-Engine Quantitative Portfolio Engine Cross-sectional multi-factor alphas, CVXPY convex optimization, Ledoit-Wolf covariance shrinkage, and walk-forward validation (Sharpe 2.14). Python
VenueWatch Market Microstructure Real-time liquidity and execution slippage telemetry engine validating feeds across Binance, Coinbase, Kraken, and OKX (3.4M events/s). Python
ApexFlow Crypto ETP Execution Low-latency crypto ETP market-making and cross-exchange arbitrage research platform. Python / C++
IMC-Prosperity-4-Backtester Backtesting Infrastructure Pure-Python backtester with Sharpe/Calmar metrics, drawdown analysis, and Google Colab execution support. Python

📈 GitHub Telemetry & Stats

GitHub Stats GitHub Streak
Top Languages
Engineering resilient agentic intelligence and high-throughput infrastructure · Aayaann Kausar

Pinned Loading

  1. Autonomous-AI-Co-Scientist Autonomous-AI-Co-Scientist Public

    Autonomous AI research platform implementing and extending Google's AI Co-Scientist with modular multi-agent orchestration, retrieval-augmented reasoning, long-term memory, automated evaluation, be…

    Python 2

  2. IMC-Prosperity-4-Backtester IMC-Prosperity-4-Backtester Public

    ⚡ Pure-Python Backtester for IMC Prosperity 4 — Python engine with built-in PnL charts, drawdown analysis, Sharpe/Calmar metrics, and Google Colab support. Clone → overwrite trader → run.

    Python 1

  3. ApexFlow ApexFlow Public

    Low-Latency Crypto ETP Market-Making & Cross-Exchange Execution Research Platform

    Python 1

  4. Multi-Alpha-Engineering-Convex-Optimization-Engine Multi-Alpha-Engineering-Convex-Optimization-Engine Public

    AlphaStack is a professional-grade quantitative research framework designed to discover, combine, and optimize multi-alpha signals into market-resilient portfolios

    Jupyter Notebook 1

  5. VenueWatch VenueWatch Public

    VenueWatch normalizes, measures, and statistically validates real-time order book liquidity, execution slippage, and infrastructure reliability across Binance, Coinbase, Kraken, and OKX.

    Python 1