Yevoo Justice
Selasi

TenetXAI — AI Agents Governance

• Extended destructive-operation detection to NoSQL and cloud-native databases, raising coverage from 0% to about 90%.
• Closed a blind spot in the critical-service-kill policy, enabling it to detect and block OS-level termination of 20 database services that previously went undetected.
• Added Google Antigravity as a supported coding agent across the Python backend, with no changes to existing agent integrations.

Bihar Krishi — Multi-Agent Chat Platform

• Reduced audio processing latency by 70% by migrating text-to-speech processing to the frontend.
• Cut LangGraph routing errors by 90% by redesigning the router to reliably parse LLM responses and recover when they came back malformed.
• Built a metadata-driven UI Actions system that lets the backend trigger buttons and page navigation in the chat UI, removing hardcoded frontend logic.
• Stopped users being wrongly logged out by adding token and profile checks on every query and stricter OTP detection.
• Built an embeddable chatbot widget that third-party sites can add without changing their tech stack, deployed across government platforms (Pucar).

Agent Framework — AI Agent Orchestration Platform (LangGraph)

• Built a shared context system that lets workflow steps pass data downstream, surviving long chains and message summarization.
• Implemented Cedar-based access control for the agent platform, restricting which users and roles can view or use tools, models, and workflows per tenant, with a no-code policy builder UI.
• Built full-stack sliding-window memory summarization for agents, preventing production context overflows.
• Diagnosed and fixed a critical shared-context bug where LangChain MCP tool outputs weren't reaching downstream agent nodes, tracing the root cause to how tool message content is structured.

AI-Assisted Project Manager — Bill & Melinda Gates Foundation

• Built Claude Skills that read each grantee's project phases from the program officer's tracker and generate tailored monthly update requests with the right sections and attachments.
• Automated reply tracking with a twice-daily routine that flags grantee changes for program officer approval before the tracker is updated.
• Used Microsoft Power Automate to send update emails and save reply attachments to OneDrive, keeping the tracker in sync.

PUCAR & Chat SDK

• Fixed connection failures between LangGraph agents and the UI Action MCP server, restoring production reliability.
• Migrated the React chat app to a framework-agnostic Vanilla JS SDK that third-party sites can embed.
• Built a UI Actions SDK that renders backend-driven action instructions in the chat.
• Integrated Bhashini STT/TTS with a Google Speech API fallback for resilient multilingual audio.

• Co-founded and built a React Native (Expo Router) shopping app with real-time buyer-vendor chat (Supabase), media-rich messaging (Cloudinary), and infinite video scrolling.
• Designed and maintained a scalable 100+ endpoint RESTful API (Express.js, MongoDB) powering all mobile/admin functionality.
• Built a URL-shortener and SSR-powered deep-link system (Next.js), plus a subscription-based push notification system.
• Built a multi-vendor order system with split orders, real-time status updates, and delivery-fee calculation via the Google Distance Matrix API.

• Built the full platform end-to-end — student, coach, and employer portals — from architecture through deployment, while teaching full-stack development to 2 cohorts as Software Engineering Instructor.
• Developed a student dashboard with lesson/project scores, career advisories, and progress charts (Highcharts), plus real-time chat/video (VideoSDK.live) for coach sessions and interviews.
• Created Talent Passport System and CareerPilot, analyzing CVs and Talent Passport data into a React Flow diagram of student strengths.
• Implemented a Job Tracker System using Python/Selenium to scrape job listings and match them to students based on their strengths.
• Delivered the coach and employer platforms — expertise-based coach registration with scheduling, and employer onboarding with project posting, candidate shortlisting, and interview tools.

• Built a Next.js/React rental platform with PostgreSQL for farmers to find and rent equipment nearby.
• Built a USSD server on AWS EC2 so non-smartphone users could place orders.
• Implemented real-time chat between farmers and equipment owners.
• Owned the backend, database, and deployment (cPanel + rsync) end-to-end.

• Built REST APIs (Node.js/Express) for course registration, auth, and user management with JWT sessions.
• Designed and optimized PostgreSQL schemas for users, courses, and transactions.
• Worked with frontend devs on API integration and consistent error handling.

• Built a multi-agent research pipeline in LangGraph that generates AI reporters for a topic, interviews an expert agent grounded in live Tavily web search, and compiles a cited final report.
• Ran reporter interviews in parallel using LangGraph's Send API across three composed subgraphs, merging results with state reducers.
• Added a human-in-the-loop checkpoint so users can approve or give feedback on the reporter panel before the run continues.
• Kept expert answers grounded by restricting them to retrieved search results, with source citations for each claim.