Ram Pakanayev
FrontEnd Developer
AI Engineer specializing in production agentic AI systems, Model Context Protocol (MCP) tooling, and multi-model LLM orchestration. Ships and hardens AI assistants from consumer scale to secure, self-hosted enterprise deployments. Full-stack across agentic backends, real-time streaming, and React frontends. Deep focus on reliability, LLM evaluation, and observability.
AI Engineering Experience
AI Engineer
- Contributing to Angie, Elementor's MCP-based agentic AI assistant embedded in the WordPress and Elementor editors — Elementor powers ~13% of all websites (millions of sites) — running in production at consumer scale (80,000+ active installations).
- Built an asynchronous MCP post-evaluation tool that scores agent responses with an LLM-as-judge across multiple quality dimensions, pinpointing where the assistant most often falls short so engineering effort lands where it matters.
- Built self-validation, an MCP tool that intercepts each draft response before it reaches the user, scores it against the tool-call trace, and returns the original when valid or a corrected one when the agent hallucinated — shipped as a reusable, feature-flagged module.
- Shipped Angie Revisions, a zero-data-loss rollback system giving users near-instant undo, redo, and restore of any prior version of AI-generated code via a non-destructive, append-only store.
- Hardened reliability by eliminating a recurring authentication error and cutting duplicate error-telemetry noise — single-flighting token refresh, consolidating error groups, and capping consecutive tool-failure loops.
- Reduced per-session code-model cost and latency at statistically equivalent task success by running model A/B experiments through a feature-gated experimentation framework.
AI Engineer
- Improved retrieval accuracy over vector-only RAG across multiple enterprise codebases by engineering CodeAI, a secure self-hosted Hybrid GraphRAG platform (Neo4j + pgvector) that models AST nodes, inheritance, and call graphs, with AWS Cognito RBAC keeping proprietary code inside the customer VPC.
- Built a legacy-SQL modernization agent (refactor, debug, analyze, generate) with strong intent-classification accuracy via AWS Bedrock, plus an autonomous Plan/Generate/Review LangGraph workflow across specialized nodes with DynamoDB persistence for multi-turn conversations.
- Improved medical-terminology accuracy for enterprise clients (hospitals, universities) by building a serverless AWS transcription platform with custom Hebrew handling, domain vocabularies, near-zero-latency WebSocket streaming, and Bedrock summarization.
- Architected a real-time Hebrew voice-to-voice system with natural turn-taking — playback begins before response completion — by pairing AWS Transcribe Streaming with a containerized Hebrew TTS on SageMaker and phrase-level chunking, closing a gap in native Hebrew cloud services.
- Built an on-demand database provisioning service with the Docker SDK, spinning up isolated, per-project environments for each engagement.
- Integrated a multi-provider Git ecosystem (GitHub, GitLab, Bitbucket, Azure DevOps) with OAuth 2.0 and HMAC-SHA256 webhook verification.
Full Stack AI Engineer
- Took MeComics, an AI-driven B2C SaaS for personalized comics, from concept to production in under 3 months as the primary developer and technical lead.
- Sharply cut comic-creation time and lifted user retention by architecting a multimodal Text2Comic pipeline (GPT-4 + Stable Diffusion) with character persistence, regeneration, and reference-image conditioning.
- Implemented reference-image conditioning with ControlNet to keep characters consistent across pages and regenerations.
- Served as technical lead across the full platform — 13 major system components and 50+ features — aligning architecture decisions with business goals.
- Built the end-to-end B2C SaaS stack — React and Node.js front to back, Stripe billing, and MongoDB — owning both product and infrastructure.
Work & Projects
Production Work
Systems running in production from my roles at Elementor and Elad Software Systems. Proprietary client names and UI are anonymized where the work is confidential.
Angie · ElementorAngie — Agentic AI for WordPress
Elementor's MCP-based agentic AI assistant, embedded in the WordPress and Elementor editors — a platform powering ~13% of all websites. As part of the Angie team, I contribute MCP tooling: response self-validation, zero-data-loss revisions, and asynchronous LLM-as-judge evaluation.
- MCP tool & server design
- Self-validation + LLM-as-judge evaluation
- In production · 80,000+ active installations
Enterprise Code Intelligence Platform
Secure, self-hosted GenAI platform designed for strict data sovereignty. Implemented a dual-backend architecture (GraphRAG + Vector) to power autonomous agentic workflows for legacy code modernization.
- Dual-backend GraphRAG + Vector architecture
- Autonomous agentic workflows
- Secure self-hosted enterprise solution
Hebrew Voice Agent
Advanced voice-to-voice agentic system powered by Strands. Features real-time orchestration of STT, LLM, and TTS. Deployed multiple fine-tuned Hebrew TTS models from Hugging Face to SageMaker (BYOC) to evaluate which performed best for production use.
- Powered by Strands Agent Framework
- SageMaker BYOC deployment for Hebrew TTS models
- Real-time voice-to-voice orchestration
Cloud-Native Audio Transcription & Analysis
Serverless audio processing platform handling concurrent workloads with near-zero latency and multilingual AI summarization. Built with AWS Transcribe, Bedrock, and FastAPI.
- Near-zero latency transcription
- Multilingual AI summarization
- Serverless concurrent processing
MeComics AI Platform
AI-driven B2C SaaS platform for personalized comics, featuring a multimodal Text2ComicPage pipeline combining GPT-4 and Stable Diffusion. Delivered MVP in under 3 months.
- Multimodal Text2ComicPage pipeline
- MVP delivered in <3 months
- Personalized comic generation
Personal Projects
Built on my own — my ideas, end to end.
Nadla — AI Real-Estate Agent
A conversational AI agent that finds your exact apartment from a natural-language Hebrew chat (“3 rooms in Be'er Yaakov under ₪2M with a balcony and parking”). Built on LangGraph with 21 tools and dynamic per-turn tool selection, cross-thread memory of your preferences and past rejections, and a self-critiquing reflection loop that revises its own recommendations — model-agnostic across Anthropic, OpenAI, and local models.
- Conversational agent that finds your exact apartment
- LangGraph · 21 tools · dynamic per-turn tool filtering
- Cross-thread memory + self-critiquing reflection (LLM-as-judge)
Consensus Chat - Multi-Agent AI Debate Platform
Real-time multi-agent AI debate orchestrator featuring sequential agent processing, live influence tracking via meta-LLM analysis, and force-directed graph visualization of idea propagation across Claude, GPT-4, Gemini, Grok, and LLaMA.
- Sequential multi-agent debate orchestration
- Real-time influence analysis with meta-LLM
- Live D3 force-directed network visualization
SilentCV — AI Job-Intelligence Platform
A four-agent system that automates the tech job hunt end to end: a Scout agent discovers and dedups roles, an Architect matches them with Hybrid RAG and drafts CVs and cover letters through an “honesty filter,” an Analyst classifies email replies via the Gmail API, and a Strategist mines rejection patterns for insights.
- Four LangGraph agents: Scout, Architect, Analyst, Strategist
- Hybrid RAG matching (pgvector cosine + PostgreSQL BM25)
- Verified-facts “honesty filter” on all generated content
CVE → WAF-Rule AI Agent
A self-correcting agent that turns a CVE into a validated AWS WAF rule. A Temporal pipeline enriches it with real signals (NVD details, CVSS, CWEs, plus CISA KEV exploit data), then a LangGraph agent drafts, validates, and revises the rule with a rationale — automating the first 80% of manual edge defense.
- Self-correcting draft → validate → revise agent loop
- Temporal enrichment from real public CVE sources
- Outputs a validated WAF rule with rationale
Fake News Generator
A satirical news app that scrapes real headlines from BBC, CNN, and ESPN, sends them to GPT-4o-mini to spin humorous fakes, and streams the results into a polished dashboard over Server-Sent Events — with live progress and AI category classification.
- Real-time SSE streaming UI with live progress
- Multi-source RSS scraping (strategy pattern)
- AI-generated titles + category classification
AI Engineering Skills
Agentic & Generative AI
AI Ops & Observability
AI Coding Tools
Cloud & Data
Backend & Frontend
Languages, Security & DevOps
Education
Bachelor of Science in Computer Science
Comprehensive education in computer science fundamentals, algorithms, data structures, software engineering principles, and emerging technologies. Focus on practical application and real-world problem-solving.
Key Areas of Study
Notable Achievement
Graduated with strong focus on emerging AI technologies and practical software development, providing the foundation for advanced work in machine learning and artificial intelligence.
Tools & Writing
Frequently Asked Questions
An illustrative, terminal-style Q&A — answers written by Ram, not generated live by these models.