AI builtfor impact.

AI Concierge that books and reschedules in plain language, plus a no-show prediction model that scores risky appointments and triggers confirmation reminders.

Web + iOS + Android apps on a shared Next.js / Prisma / Postgres core, Stripe Connect for clinic payouts, Redis-backed slot management, an AI Concierge that handles reschedules and FAQs, and a no-show predictor that surfaces risk directly in the clinic dashboard.

Modeled clinics, services, staff, and slots in a multi-tenant schema with subdomain routing; layered Capacitor on top of the same React surface so mobile reuses the booking flow; trained the no-show classifier on historical booking + cancellation patterns.

AI Ops Agent reachable over MCP that answers natural-language questions, settles driver bonuses, pulls live GPS, and runs anomaly detection on incoming earnings.

A Next.js 16 dashboard with multi-tenant RBAC, an EC2 Playwright service that scrapes platform CSVs nightly, an EV fleet telemetry deployment that streams vehicle WebSocket data into the same database, and an embedded AI Agent reachable from inside the app.

Wired a JSON-RPC MCP server so the scraper, the telemetry forwarder, and the AI Agent all share one typed contract, so the agent can answer 'who hasn't checked in?', settle a driver bonus, or pull live GPS without any direct DB access. The same MCP server powers earnings anomaly detection.

AI Recruiter that parses CVs into structured fields, scores candidates against the job, and ranks the shortlist; a separate AI Writer drafts job descriptions and career-site blog posts from a short brief.

Next.js 14 ATS on Prisma + Postgres, NextAuth for admins, magic-link flows for candidates, Resend for transactional email, Google Calendar two-way sync, an AI Recruiter pipeline that extracts skills and experience into structured fields, and a job-fit scorer that ranks applicants per role.

Split the codebase into `admin/` and `prospect/` component domains so one data model serves both surfaces; the AI pipeline runs on every CV upload, writes scores back to the application, and surfaces a ranked shortlist directly in the recruiter inbox. The AI Writer reuses the same JD context to keep job posts and blog content consistent.

AI Triage Agent that classifies threads against a label taxonomy, routes them to the right queue, and drafts a reply the operator can edit and send in one click.

Next.js 16 + Supabase app with an AI-driven classification and reply-drafting pipeline, Gmail webhooks served through Supabase Edge Functions, and a Postgres-changes realtime hook that keeps the queue UI live without manual refresh.

Used Supabase RLS for multi-operator access, kept queries composable by passing the client as the first argument, and centralized Gmail labels, statuses, and colors in a single constants module so the UI and the AI agree on state.

Hierarchical AI Agent system: a Campaign Orchestrator supervises Strategy, Content, Execution, and Analytics teams, each an AI Agent with its own tools, prompts, and structured outputs.

Next.js 14 frontend, NestJS GraphQL backend, and a Python/LangGraph agent service split into Strategy, Content, Execution, and Analytics teams under a Campaign Orchestrator supervisor, where each team is an autonomous AI Agent with its own tools, prompts, and structured outputs.

Connected services over Kafka so any team can subscribe to campaign state changes; modeled audience and campaign relationships in Neo4j alongside Postgres for transactional data, with Composio giving every agent a uniform tool surface across ad platforms, CRMs, and social channels.

AI Debate Coach that scores arguments across 7 axes, generates a steelman of the user's own position, asks Socratic counter-questions, and replies in a tone the user picks.

Browser audio capture piped through a speech-to-text service for transcription, an AI Coach for analysis and steelman generation, and a tone-adapted text-to-speech voice for spoken feedback.

Designed the loop to feel instant: keyboard shortcuts (Space / Esc / Ctrl+D), a demo mode with cached results for live presentations, and auto-save so practice sessions never disappear mid-flow.

A trio of AI agents shapes the candidate journey end-to-end: an AI Onboarding Concierge guides candidates through agreements and payment plans in their native language, an AI Document Reviewer pre-screens passports and evaluation files before they reach a human admin, and a Smart Reminder Agent scores at-risk installments and stalled lessons to time the next nudge per channel and locale.

A Next.js 15 App Router app with a candidate portal (i18n + RTL) and an English-only, frame-locked admin portal. Custom JWT sessions via jose with 30-day sliding cookies, four sign-in modes (OTP, magic link, password, SSO), and multi-tenant billing-firm logic with per-firm invoice numbering, director signatures, and counter-signed PDF agreements. Bidirectional elec.work integration with HMAC-SHA256 signed webhooks both directions, idempotent backfill, and exponential-backoff retry. Field-level AES-256-GCM encryption on PII and 7-year audit-log retention on every admin mutation.

Built end-to-end by a structured Claude Code agent pipeline: 18 specialized subagents (architect, prisma-schema, payments, auth, elecwork-integration, ux-ui, qa-unit, qa-e2e, code-review, deploy, docs) orchestrated through a 6-stage YAML ledger (plan, implement, verify, ship-dev, ship-prod) with a code-review gate after every commit and mandatory documentation updates as the final step of each task. Sprints close only when every task is shipped to prod or formally rejected. The pipeline is the artifact: 39 production tasks shipped across 7 sprints with zero context loss between agent handoffs.