Hi, I'm Kush Ise

Built a competitive intelligence platform as an OpenClaw plugin that equips sales teams with AI-generated battle cards before every call. Features a parallel five-actor Apify scraping pipeline (competitor pricing, Google mentions, LinkedIn signals, G2 reviews), Redis-backed semantic memory layer with agent-memory-server on Redis Cloud, and a semantic cache returning results in under 100ms for repeat queries. On cache miss, triggers a four-hop grounded retrieval chain through Contextual AI with full citation chains making hallucination structurally impossible. Features Civic Auth JWKS-based JWT verification with RBAC, and an interactive 3D dashboard with React 19, Three.js, and five AI sales characters.

Built an automated security scanner and auto-remediation tool for AI coding agents. AgentTrace provisions a live OpenClaw agent inside a Blaxel cloud sandbox, attacks it with 10 security scenarios across 6 threat categories (prompt injection, sandbox escape, credential theft, persistence, evasion, config exploit), scores each attack using Claude Sonnet as LLM-as-a-judge, auto-patches the agent configuration to fix discovered vulnerabilities, and re-scans to verify the fixes. Delivers a measurable before/after security posture improvement (e.g., Grade C to Grade B). All 40 traces logged to Opik for full observability.

Built an AI-powered intelligence marketplace that autonomously buys and sells real-time world intelligence using the Nevermined x402 payment protocol. Sells geopolitical briefings, market data, cyber threat intel, and deep research to buyers for financial and investment planning -- all paid per request on-chain. Autonomously discovers and purchases intelligence from other AI agents on the marketplace and resells in real time. Features a live interactive command center dashboard with military conflict tracking, financial markets, aviation, maritime, oil and gas pipelines, and 20+ real-time data categories. Built on Strands Agent (Claude Sonnet), Apify, Exa, and Ability.ai Trinity with full A2A agent interoperability.

Built a multi-agent autonomous risk intelligence system that researches, analyzes, and scores vendor risk across 73+ regulatory checkpoints (GDPR, CCPA, ISO 27001, NIS2, EU AI Act, DPDP) in minutes instead of weeks. Four specialized AI agents (Procurement, Compliance, Cybersecurity, Analyst) simulate real-world vendor review workflows. Features AI-powered document analysis, composite risk scoring with dynamic weighting, human escalation for ambiguous decisions, and audit-ready reports with full reasoning transparency. Uses temperature tuning for balanced flexible reasoning and deterministic scoring.

Built an OpenClaw plugin that serves a full browser dashboard directly from the gateway -- no separate server, no config file editing. Features a live action feed that translates every agent tool call into plain English in real time, an approval rail that pauses risky actions (bash commands, file deletes, emails) and waits for user confirmation before executing, and a visual Slack setup wizard that connects OpenClaw to Slack in 4 browser steps with zero JSON editing. Uses clawg-ui from CopilotKit as the AG-UI layer to bridge the OpenClaw gateway to the React frontend, streaming agent events in real time. Built in 5 hours.

ClawGuard is a security-first filtering server that sits between external inputs (email, webhooks) and LLM agents, preventing prompt injection, secret exfiltration, and malicious content from ever reaching the AI. Connects to Gmail via GCP Pub/Sub for real-time email ingestion, runs a deterministic sanitization pipeline that strips HTML/scripts, detects 10+ injection patterns, redacts secrets, removes hidden unicode characters, and scores every message with a 0-100 risk rating. Exposes a REST API for agent skills to query sanitized emails on demand, and includes a live dashboard showing raw vs. sanitized content side-by-side with color-coded risk highlighting.

A native macOS application that reimagines Spotlight as an AI-powered personal search engine. Built with SwiftUI and a FastAPI backend, it combines semantic vector search, temporal file discovery, browser history lookup, and live web search into a single unified interface. Runs MLX embeddings (bge-small) directly on Apple Silicon to index local files -- documents, code, PDFs -- enabling natural language queries. Claude synthesizes RAG-powered answers grounded in actual files and web results with full source attribution. Features an integrated file browser with Quick Look previews. Private-first -- all embeddings and indexing happen locally on-device. Built in a single day for Better Hack at YC.

Built a transparent AI agent evaluation platform that lets users run live benchmarks powered by Gemini, conduct voice evaluations via Pica with live audio streaming and cleanup, test real-time web intelligence, perform tool verification, and submit community reviews. Features an agent submission system, benchmark leaderboard, and side-by-side comparisons. Helps users make informed decisions about AI agents through standardized, reproducible evaluations across multiple dimensions.

Built a persistent desktop-native voice agent with an ambient companion that maintains always-on context across your entire workspace. A spatially-aware AI copilot that floats across windows and desktops, observing screen state and responding through real-time speech-to-speech. Uses Gemini Live for native S2S which collapses the traditional STT-LLM-TTS cascade into a single low-latency service. Routes through Daily's SFU for WebRTC transport with Pipecat orchestrating the pipeline. Supports calendar management, note-taking, and contextual Q&A.

Built an AI-powered assistant for querying 160,000+ San Francisco government datasets from data.sfgov.org, running fully local on Dell Pro Max GB10 units with NVIDIA DGX Spark. Features natural language queries, semantic search across departments, cross-agency data alignment to surface contradictions and reduce ambiguity, voice input via ASR, SQL generation from plain English, and RAG for grounded, traceable answers. Helps support better decisions, transparency, and public trust by unifying siloed government data into a single AI interface.

Built an AI-powered Reddit marketing automation tool that creates authentic, engagement-driven content at scale. Features a sophisticated planning algorithm with hard constraints to prevent overposting (max 2-3 posts per subreddit/week), semantic deduplication to catch similar topics, LLM-based quality scoring with automatic regeneration, and strategic distribution that rotates personas, subreddits, and timing to avoid detection patterns. The system thinks strategically about distribution, tracks used topics across weeks, validates authenticity, and regenerates low-quality posts up to 3 times with improvement guidance.

Built an AI-powered HR automation system with Next.js and Vercel AI Gateway that streamlines HR workflows end-to-end. The app features task management, AI-generated HR documents (offer letters, job descriptions, onboarding guides), and automated email delivery via Resend. Includes a kanban-style task board for visibility and status tracking, helping People Ops teams reduce repetitive work and produce consistent, professional documents in seconds.

Built multimodal AI assistant developed for clinical researchers and radiologists. It integrates a custom Model Context Protocol (MCP) to coordinate multiple large language models and retrieval tools, enabling real-time access to web and PubMed data. The system uses a Retrieval-Augmented Generation (RAG) pipeline to provide domain specific answers to medical queries and generate preliminary radiology reports from image inputs. Built with Python and Streamlit, offers an intuitive interface for both text and image-based interactions, streamlining medical research and diagnostic workflows.

Zuno is an intelligent, multimodal shopping advisor with a Streamlit UI and a Fast API gateway backed by a LangChain/Gemini multi agent stack. It accepts text or image inputs (with EasyOCR for query extraction), uses SerpAPI to retrieve products, similar images, and video reviews, and applies a Comparison Agent to weight price vs. rating. An Orchestrator Agent coordinates tools and delivers a concise, human-readable recommendation via a single clean API.

Built an AI-powered web app to assess credit risk from debit transaction history. Applied unsupervised clustering to generate pseudo-labels and trained a neural network for multi-class risk prediction. Integrated Google Gemini API for LLM-generated insights and financial recommendations. Secured PDF report downloads with Auth0.

Developed an autonomous driving system for a solar electric vehicle, including lane detection, IoT integration, and dynamic path planning. The project was highlighted in EVreporter e-magazine and raised 1.2 lacs INR.

Developed an unsupervised machine learning algorithm using CNN to count and track trees in high-resolution satellite and drone images. Achieved 98% accuracy, making it valuable for the palm and agriculture industry.

Developed an interpreter in Python to evaluate basic math calculations. The project involved lexer and parser components, providing insights into how computers process human-readable text.

Developed a project to classify X-ray scans as healthy, pneumonic, or COVID positive. Used ResNet18 network pre-trained on ImageNet and achieved 98% accuracy in COVID detection.

Built an interactive shopping cart UI using React, allowing users to browse and add/remove shoes from the cart seamlessly. Integrated cart functionality, product filtering, and checkout processes, enhancing user experience.

An AI-powered portfolio chatbot that showcases my skills, projects, and resume, using the Gemini API.