John Dominion

<p>Built a human-in-the-loop email-approval agent system in Go, connecting an AI orchestrator, email service, and approval service through a GraphQL API, with every AI-drafted email routed through mandatory review before sending.<br>
Enforced a two-stage execution pipeline that blocks the LLM from triggering sends directly.<br>
Hardened database mutation paths so sensitive email operations require authenticated human-user authorization and cannot be triggered by service tokens.</p>

<p>Built a real-time voice AI system on Google Gemini Native Audio over WebSocket, streaming 16kHz PCM audio and generating live JSON transcripts alongside AI speech responses.<br>
Cut conversational turn-taking latency by 400–600ms by tuning Voice Activity Detection sensitivity and silence-window calibration.<br>
Co-built a multi-agent career platform with LangGraph and Gemini 2.5 Flash covering skill-gap analysis and personalized learning roadmaps.<br>
Built unified Redis session memory so context carries over seamlessly between text and voice channels.</p>

<p>Built a RAG-based legal chatbot with a dual-index Pinecone setup (sparse keyword + dense semantic) supporting keyword search, semantic retrieval, and reranking across thousands of legal documents.<br>
Built an OCR pipeline that extracts and translates scanned Arabic PDFs into searchable text using a vision model.<br>
Built an async ingestion pipeline for PDF and DOCX files that enriches each chunk with metadata to improve retrieval precision.<br>
Added concurrent user session memory, containerized services with Docker, and wrote onboarding documentation for the team.</p>

<p>Improved course-generation speed with async module processing and GPU-accelerated inference inside an AI-powered LMS that generates multimedia courses using LLMs and RAG.<br>
Wrote parser functions that validate and auto-correct LLM outputs to keep generated course content well-structured and consistent.</p>

<p>Led development of a document-intelligence platform on GPT-4 and Azure OpenAI that ingests PDFs and DOCX files and supports natural-language querying, summarization, and content generation.<br>
Built the RAG backend using Azure OpenAI embeddings, Azure Cognitive Search as the vector layer, and Redis for session state.</p>

<p>Built a RAG-powered support chatbot that cut response time by 65% and reduced human workload by 70%.<br>
Built a CV-ranking and summarization tool using Gemini and cosine similarity, improving match accuracy by 40% and review speed by 80%.<br>
Built a visual fashion recommender (ResNeXt50 + KNN) that lifted e-commerce engagement by 20%.<br>
Built a RAG-powered dental-health chatbot that cut consultation overhead by 35%.<br>
Built a real-time weapon-detection system using YOLO and Flask for live surveillance analytics.<br>
Built an LSTM model for gold-price forecasting with automated GitHub Actions retraining.<br>
Built an AI data-analysis app with Pandas AI and OpenAI that improved reporting speed by 40%.<br>
Built an automated hackathon code evaluator that cut grading time by 60%.</p>

<p>Built an AI content-creation pipeline for multilingual localization and interactive storytelling using OpenAI and Groq Audio.<br>
Built a data-visualization tool integrating Groq models with Matplotlib that cut manual effort by 60%.<br>
Created courses and Jupyter notebooks for the developer community.</p>

<p>Built a real-time voice-interviewing agent on Gemini Native Audio over WebSocket, with Redis session persistence and transcript recovery so interviews survive connection issues.</p><p> Built a semantic resume-matching engine using Gemini embeddings and cosine similarity to align candidates with job requirements.</p><p> Built a standalone microservice that reads interview transcripts from Redis and generates structured scorecards covering communication, technical depth, and role fit.</p><p> Co-built an agentic project-scoping tool that turns a user brief into a structured PRD, then matches candidates via Pinecone Serverless with 3072-dimension Gemini embeddings at near O(1) retrieval speed.</p><p> Built async indexing pipelines with Redis change detection that skip unchanged candidate files, significantly cutting embedding and vector-upsert costs.</p>

<p>Built a reusable web-scraping framework that converts successful sessions into config templates, letting future runs replay with zero LLM calls.<br>
Built an eBay scraping agent directly on Browser-Use's low-level APIs, avoiding nested orchestration overhead and cutting memory use.<br>
Built a recursive DOM parser with five levels of parent traversal, keeping product-data extraction reliable even when site layouts change.</p>

<p>Built a multi-modal auto-grading agent that evaluates handwriting, images, audio, video, and text responses, including math and computation-heavy subjects most grading tools skip.<br>
Rebuilt the question-generation engine to run 5x faster at roughly 1/30th the cost per run, while improving output quality.<br>
Built a 12-stage asynchronous plagiarism and authorship-detection pipeline connecting SerpAPI web discovery, page scraping, Gemini embeddings, semantic similarity scoring, citation validation, and stylometric analysis.<br>
Built the Examurai grading platform end-to-end, including async OCR extraction, ARQ worker orchestration, webhook delivery, and Redis idempotency controls that prevent duplicate LLM calls at scale.<br>
Designed a personalized development-planning agent for educators using a LangGraph multi-agent workflow, delivered in partnership with the Universal Service Provision Fund (USPF).<br>
Separated the AI layer into its own FastAPI service and built multi-tier Redis caching with MongoDB indexing to cut external API costs across SerpAPI calls, scraping, and embeddings.</p>

<p>Built a multi-agent career-counseling platform with stateful LangGraph workflows that runs adaptive interviews and generates personalized career pathways and development plans.<br>
Built a self-healing layer for structured LLM outputs that detects malformed responses and automatically reconstructs a valid format.<br>
Built a retrieval pipeline that validates every learning-resource link in real time before it reaches the user, eliminating hallucinated or broken recommendations.<br>
Set up Redis-backed session management with multi-layer auth and proxy-aware IP resolution for secure cloud deployment.</p>