AI Engineer · UAE

Athira
Anil Kumar

Building production-ready AI systems — from multi-agent RAG pipelines to LLM security frameworks. MSc Cyber Security with Distinction. IEEE-published researcher.

Generative AI Agentic AI · LangGraph LLM Security AI Red Teaming RAG · FAISS · BM25 FastAPI · Docker
View My Work Get In Touch
23+
AI Projects
1
IEEE Paper
98%
Model Accuracy
About

Building AI that
actually works

I'm an AI Engineer based in Sharjah, UAE, specialising in Generative AI, Agentic AI systems, and LLM Security. I build things end-to-end — from architecture to deployment.

My background spans transformer-based NLP, multi-agent orchestration with LangGraph, RAG pipelines with hybrid retrieval, and AI red teaming. I've published research at IEEE on AI-generated phishing detection and hold an MSc in Cyber Security with Distinction.

What drives me is the intersection of capable AI and secure AI — systems that are intelligent and robust.

MSc Cyber Security — Distinction University of West London, 2025
IEEE Published Researcher AI-Generated Phishing Detection, 2025
UAE Resident — Available Immediately Open to on-site, hybrid, or remote roles
Current Role
GenAI Security Testing Developer
Cosmic Info Solution · May 2026 – Present
AI Red Teaming · LLM Security
RAGAs Evaluation
Faithfulness: 0.91 · Relevancy: 0.88
Production RAG system with CI/CD quality gates
Agentic RAG · LangGraph · Docker
Reinforcement Learning
PPO · Q-Learning · DDA
Shipped RL agents in real game projects
Gymnasium · Stable Baselines3 · Godot
Skills

The stack

Tools and technologies I use to build production AI systems.

Generative AI & LLMs
RAGGraph-RAGLangGraphLangChain Agentic AIPrompt Engineering Corrective RAGEmbeddings FAISSBM25Ollama RAGAsCross-Encoder Reranking
AI Security & Red Teaming
Prompt Injection TestingJailbreak Testing System Prompt ExtractionAI Red Teaming LLM Security EvaluationAdversarial Prompting Severity Classification
Machine Learning & NLP
TransformersDistilBERT Scikit-LearnSHAPLIME TF-IDF SearchFeature EngineeringNLP Pipelines
Reinforcement Learning
PPOQ-Learning Stable Baselines3Gymnasium Dynamic Difficulty Adjustment
Backend & DevOps
FastAPIDocker GitHub ActionsCI/CD Azure Container AppsSSE Streaming REST APIsn8n
Languages & Tools
PythonJavaScript C / C++GDScript GitLinux Godot 4
Projects

Things I've built

Production AI systems, automation pipelines, security tools, and RL-powered games.

Security GenAI
⚔️ 11 attack categories · full campaigns in seconds · CI regression gate
AgentShield — AI Agent Red-Team Platform
Autonomous red-team platform that builds an intentionally vulnerable multi-agent target, then attacks it with a closed-loop agentic engine — Attack → Evidence → Impact → Remediation → Regression Test. Covers prompt injection, RAG/memory/tool poisoning, MCP attacks, and agent-to-agent exploits, with a live security-scorecard dashboard and a GitHub Action that gates CI on regressions.
FastAPILangGraph FAISSDockerGitHub Actions
Security
🔐 p50 243ms · p95 563ms @ 69 req/s throughput
AegisAI — Zero-Trust Runtime Security Gateway
Runtime security layer between AI agents and their resources — verified agent identity, policy-as-code access control via OPA/Rego, prompt-injection and behavioral anomaly detection, and a human-approval escalation path for high-risk actions. Full observability stack with OpenTelemetry, Prometheus, and Grafana.
FastAPIOPA/Rego RedisPostgreSQLGrafana
Automation GenAI
🔗 HMAC-signed webhooks · AI agent requires approval before writes
Full-Stack CRM Platform
Production CRM with contact/company management, a drag-and-drop deal pipeline, role-based access, and REST API with HMAC-signed, retry-safe webhooks. Optional AI layer adds deal-insight summarization and an agentic assistant that requires explicit user approval before any data modification.
Next.jsTypeScript PostgreSQLPrismaClaude API
Automation GenAI
📚 Double-entry accounting · 7 role-based access tiers
Trading ERP Platform
Full-stack ERP for trading companies unifying sales/CRM, purchasing, multi-warehouse inventory with weighted-average costing, and double-entry accounting with bank reconciliation. Includes HR/payroll and an optional Claude-powered copilot for business queries and bill data extraction.
Next.jsNestJS PrismaPostgreSQLClaude API
Reinforcement Learning
🤝 99.7% of theoretical max information transfer, validated across seeds
Multi-Agent Reinforcement Learning System
Cooperative, competitive, and emergent-communication multi-agent RL across three PettingZoo environments, trained with PPO. An explicit information-bottleneck architecture and mutual-information auxiliary loss let agents learn a real communication protocol — validated with multi-seed runs, not just a rising reward curve.
PettingZooStable-Baselines3 PPOPython
Security GenAI
🛡️ 100% recall · 0% FPR · cuts LLM compromise 90.8%→19.7%
AI Security Gateway
Security middleware between users and LLMs — detects prompt injection, jailbreaks, system-prompt leakage, PII, toxicity, and malicious tool calls in real time, then allows, flags, sanitizes, or blocks with a full audit log.
FastAPIPython DockerGitHub Actions
GenAI
🎯 Faithfulness 0.91 · Relevancy 0.88
Enterprise Agentic RAG Platform
Production-grade RAG system with a LangGraph agentic loop, hybrid BM25+FAISS retrieval, Reciprocal Rank Fusion, 2-stage cross-encoder reranking, and Corrective RAG with web search fallback. RAGAs CI/CD quality gate included.
LangGraphFAISSBM25 FastAPIDockerRAGAs
GenAI Automation
📊 p50/p95/p99 latency · cost tracking · LLM-as-judge
Production LLMOps Platform
Self-hosted observability and regression-testing platform for Claude-powered apps — tracing, cost/latency/token metrics, and an LLM-as-judge evaluation engine that gates prompt changes against statistical baselines before they ship.
FastAPIPostgreSQL ReactClaude APIDocker
GenAI
Azure Enterprise Agentic AI Platform
Enterprise-grade Agentic RAG on Azure OpenAI + Azure AI Search with Semantic Kernel routing, zero-trust identity via Entra ID and Managed Identity, and Bicep/GitHub Actions IaC with an automated quality gate on groundedness and relevance.
Semantic KernelAzure OpenAI Azure AI SearchBicepEntra ID
Security GenAI
⚡ 5-agent consensus · patches + reports in under 60s
CodeSentinel — Multi-Agent Code Security Review
Five specialized AI agents debate and reach consensus to identify vulnerabilities in code, then auto-generate secure patches and professional security reports across multiple formats.
Google ADKGemini MCPStreamlit
GenAI Automation
🧠 Zero-dependency · Offline-first · Graph-RAG
Mind MRI — Agent Second Brain
Zero-dependency, offline-first Graph-RAG memory core and REST API server for autonomous AI agents. Features vector-like TF-IDF semantic search, exponential time-decay relevance scoring, and an interactive D3.js knowledge-graph visualizer — persistent, queryable memory built for production agent pipelines.
Node.jsGraph-RAG TF-IDFD3.jsREST API
Security Automation
TrustMesh — SIM-Swap Fraud Defense
Real-time fraud defense engine using telecom CAMARA network APIs and n8n-orchestrated AI agents to detect and stop SIM-swap attacks before MENA digital payments complete, applying risk-based step-up authentication.
n8nGroq/Llama 3.3 CAMARA APIsGemini Flash
Security Automation
SOC Automation Toolkit
CLI that parses security logs (SSH auth, CSV, JSON, syslog), detects brute-force and port-scan patterns, enriches alerts via VirusTotal/AbuseIPDB/GeoIP, and generates AI-summarized reports — treating log data as untrusted input to prevent prompt injection.
PythonClaude API VirusTotalAbuseIPDB
GenAI Automation
📞 Voice AI cold-calling · booking, zero human intervention
SunPeak — AI Customer Acquisition Pipeline
End-to-end AI SDR for a solar company — a RAG-backed qualification chatbot hands off to an AI voice agent that cold-calls leads and books meetings, with a live dashboard tracking booking rate and call quality.
ClaudeVAPI QdrantFastAPIn8n
GenAI Automation
🛎️ Pilot-ready · multi-tenant
Customer Engagement AI
Automates customer intake for trade businesses (plumbing, HVAC, electrical) — answers missed calls, texts, and website chats, qualifies leads with structured questions, and books appointments directly, escalating to a human only when uncertain.
TypeScriptClaude API TwilioStripeGoogle Calendar
GenAI
AI Operations Assistant
Multi-agent business intelligence system routing questions through a Manager → Research → Analysis → Report pipeline using LangGraph StateGraph and Claude Sonnet. Real-time SSE streaming, Azure-ready Docker deployment.
LangGraphClaude Sonnet SSE StreamingAzureDocker
GenAI Security
AI Content Brief Workflow
Browser-based four-stage chained LLM pipeline — Intent Analysis → Outline → Brief Drafting → Quality Scoring. XSS-safe output, CSP restricted to api.anthropic.com, secure API key management.
JavaScriptAnthropic API Prompt ChainingCSP Security
Security Automation
Automated Cyber Threat Intelligence Feed
Zero-cost platform aggregating 9 RSS feeds (~72 articles/day) with MITRE ATT&CK auto-tagging, Groq/Llama 3 AI briefings, severity scoring, and daily automated distribution. Free REST API via GitHub Pages.
PythonGroq/Llama 3 MITRE ATT&CKGitHub Actions
Automation GenAI
AI Digital Product Pipeline
End-to-end zero-cost AI pipeline: trend research → content generation → PDF export → Gumroad publish → social scheduling → email sequences. One command, fully automated. Bundle auto-generation after 3 products in a niche.
Node.jsGemini API Gumroad APIBrevo
NLP Security
🎯 ~98% accuracy on open-source datasets
AI-Powered Phishing Email Detector
Tri-class phishing detection (legitimate / human-phishing / AI-generated) using DistilBERT with dual-threshold decision strategy. SHAP + LIME explainability for analyst trust. Basis of IEEE publication.
DistilBERTTransformers SHAPLIMESpark
Reinforcement Learning
DungeonRL — Adaptive AI Dungeon Crawler
Procedurally generated dungeon crawler with a PPO RL agent performing Dynamic Difficulty Adjustment in real time. Agent reads live telemetry every 10s and adjusts enemy stats. Same DDA concept as Resident Evil 4 and Left 4 Dead.
PPOGymnasium Stable Baselines3Pygame
Reinforcement Learning GenAI
LifeForge — AI Life Simulation Game
Text-based life simulator in Godot 4.3 with Q-Learning NPCs, Gemini Pro LLM narrative generation, dynamic economy via Ornstein-Uhlenbeck processes and Markov chains, Firebase cloud saves, and AdMob monetisation.
Q-LearningGodot 4.3 Gemini ProFirebase
Automation
AI Timetable Optimization Engine
Constraint-based scheduling engine using Google OR-Tools, with a full SaaS frontend and FastAPI backend. Live on Vercel.
PythonFastAPI OR-ToolsVercel
Research

IEEE Publication

IEEE International Conference · 2025
Detecting Human-Written and AI-Generated Phishing Emails Using DistilBERT and Explainable AI
Developed a tri-class classification framework capable of distinguishing legitimate emails, human-written phishing, and AI-generated phishing — a critical distinction as LLM-powered attacks become indistinguishable from human-crafted ones. Applied SHAP and LIME explainability techniques to build analyst trust and enable interpretable security decisions.
~98%
Accuracy
3
Classes
XAI
SHAP + LIME
DistilBERT Explainable AI NLP Phishing Detection IEEE Published
Read on IEEE Xplore →
Experience

Where I've worked

May 2026 – Present
GenAI Security Testing Developer
Cosmic Info Solution
  • Evaluating AI systems against prompt injection, jailbreak, system prompt extraction, and role confusion attacks.
  • Designing security testing methodologies to assess model robustness, safety controls, and guardrail effectiveness.
  • Generating structured security assessment reports with risk ratings and remediation recommendations.
  • Contributing to AI red teaming research and development of the Gen-AI Prompt Injection Tester.
Apr 2026 – May 2026
AI-Native Developer / Agentic Product Builder
Optima AI
  • Contributed to architecture design for an AI-powered video production platform.
  • Worked on workflow orchestration, AI model integration, and distributed compute infrastructure.
  • Designed agentic AI workflows for content generation and media processing.
Education & Certifications

Where I studied

Degree
MSc Cyber Security — Distinction
University of West London · 2025
Distinction IEEE Published

Big Data Analytics · Machine Learning for Security · Security Operations & Assurance.
Dissertation: AI-Generated Phishing Detection using DistilBERT — published at IEEE International Conference 2025.

Certification
One Million Prompters — Prompt Engineering
Dubai Crown Prince's Court Initiative, launched by HH Sheikh Hamdan bin Mohammed bin Rashid Al Maktoum · 2025
UAE Government Initiative

Completed the One Million Prompters programme training prompt engineering skills for AI systems.
View certificate →

Contact

Let's build something
intelligent

Open to AI Engineer, GenAI Engineer, LLM Engineer, and AI Security roles — on-site in UAE, hybrid, or remote. Also available for freelance AI projects.

ara.ceh24@gmail.com linkedin.com/in/athira-a-k github.com/ara-5
Available immediately · Sharjah, UAE
Download Resume →