Project case study
Cynergists AI
An agentic CRM where one AI, Cynessa, runs lead generation, follow-up, booking, campaigns, calls, and websites. It sits on a multi-tenant CRM with approval-gated actions, multi-provider LLM routing, real-time voice, and usage-based billing.
Team project with contribution from Jonas Viray, Full Stack Developer & AI Automation Engineer
Overview
My role · Full Stack AI Developer and largest individual contributor: 4,000+ merged pull requests (about two-thirds of all merged PRs) and 6,700+ commits since April 2026 across frontend, backend, AI systems, real-time voice, infrastructure, and security on the Cynergists AI product team.
Cynergists is an agentic CRM powered by Cynessa, an AI built into the platform that does the sales, marketing, and operations work a business would otherwise spread across separate tools, while customers pay for the work she performs. Cynessa is the customer-facing intelligence: she owns reasoning, planning, company context, tool choice, and approvals, and calls roughly fifty typed capabilities for CRM, calls, email, social, SEO, websites, meetings, hiring, and support. Behind her is a production system of about 1.5 million lines: a React and TypeScript web app, Supabase Postgres with row-level security and hundreds of Deno edge functions, Node and Bun services on Railway for chat, real-time voice, website-builder sandboxes, and backups, and an Expo mobile app.
engineering summary
Problem → Architecture → Contribution → Result
- Problem
- Small and mid-sized businesses run lead generation, follow-up, calls, scheduling, campaigns, websites, and operations across disconnected tools and separate AI agents. Cynergists set out to replace that with one platform and one AI that understands the business, takes real actions safely, isolates every tenant's data, and bills for the work it actually performs.
- Architecture
- React 18, TypeScript, and Vite on Vercel serve the web app; Supabase provides Postgres with row-level security, Auth, Realtime, pg_cron, and hundreds of Deno edge functions; Node and Bun services on Railway run the chat gateway, real-time voice, website-builder sandboxes, backups, and canary monitoring; and an Expo React Native app covers mobile. Cynessa sits in front of a typed capability registry with server-side execution policy, routes between Gemini, Claude, and OpenAI with fallback, and meters usage through a reserve-then-settle credit wallet. Voice runs on LiveKit, Deepgram, ElevenLabs and Kokoro text-to-speech, OpenAI Realtime, and Plivo, and Stripe handles subscriptions, refills, and partner payouts.
- Contribution
- As a Full Stack AI Developer I was the largest individual contributor: 4,000+ merged pull requests (about two-thirds of all merged PRs in the repository) and 6,700+ commits since April 2026. The work spanned the Forge AI website builder (about 930 PRs), Cynessa and the unification of specialist capabilities behind her, CRM and meetings, real-time voice and telephony, marketing AI, billing and the credit wallet, security hardening, and the CI/CD and release-safety tooling. Much of it was delivered by directing and reviewing Claude Code and Codex agents under working rules I wrote for the team, plus 59 merged pull requests on the new OpenAI Agents SDK runtime.
- Result
- A production platform of roughly 1.5 million lines (a web app, hundreds of serverless functions, more than 2,500 versioned migrations, about a dozen services, and a mobile app) where one AI can take approved actions across CRM, calls, email, social, and websites, protected by fail-closed policies, server-verified approvals, per-tenant isolation, and thousands of automated tests running in CI.
AI / integration engineering notes
- One customer-facing AI is backed by a typed registry of about fifty capabilities; each carries a risk level, an approval policy, and a readiness status, and anything that is not ready fails closed instead of improvising.
- Human approvals are verified against run state persisted in Postgres rather than a flag from the model, so forged or replayed approvals are refused; guardrails validate input, output, and actions, and the assistant cannot report an action as done unless a tool actually ran.
- Cost-aware routing across Gemini, Claude, and OpenAI with status-based fallback, and a credit wallet that reserves credits before expensive AI work and then settles or releases them.
- Real-time voice and video: LiveKit rooms, a Deepgram speech-to-text gateway, browser and phone calling over Plivo, private live AI coaching for reps on calls, and a WebSocket-to-gRPC bridge that streams NVIDIA Audio2Face-3D blendshapes to a Three.js avatar.
- The Forge website builder generates real source files with live preview, imports and edits GitHub projects behind approval gates with atomic commits and rollback, validates builds in isolated sandboxes, and publishes to custom domains.
- Autonomous agents with safety rails: an opt-in inbox agent needs an atomic claim and an idempotency key before any real send or booking, and a new agent runtime on the OpenAI Agents SDK rolls out per user and per tenant behind a kill switch with canary monitoring.
- AI-assisted delivery at scale: wrote the team's Claude Code and Codex working rules and release-safety policy, and built an autonomous pipeline where one coding agent implements, a second reviews, and a human merges.
- Production hardening across multi-tenant row-level security, database-function security audits, migration-drift detection, SHA-pinned CI with a failing-test ratchet, and encrypted nightly off-site backups.
Related service: Product Engineering & Business Systems, AI Applications & Agents
Observed product surface
- Lead generation, follow-up, appointment booking, and campaigns run by Cynessa
- Contacts, companies, pipeline, and KPI dashboards in a native CRM
- Browser and phone calling, SMS, call transcription, and live AI coaching
- Email, social, video, and SEO content tools behind one assistant
- AI-generated websites with custom domains, booking, and storefront checkout
- Usage-based billing for the work the AI performs
Public product areas
- Cynessa AI workforce
- Native CRM and Command Center
- Agent marketplace
- Calls, meetings, and booking
- AI website builder
Intended audience · Small and mid-sized businesses, sales teams, and agencies that want one platform and one AI handling marketing, sales, calls, websites, and operations
Planning a system with similar complexity?
Send the requirements, current stack, and constraints. I'll outline how I would structure the system, where the main risks are, and what an implementation path could look like.
More projects
US Fund Advisor
A production business-funding platform that combines AI analysis with account flows, financial data, payments, communications, document generation, external APIs, and multi-service infrastructure inside one end-to-end application.
Atimonan Laundry Shop
A full booking and service-operations application covering customer scheduling, order status, payments, notifications, and staff administration, backed by Express, Supabase, Redis, and production deployment on Vercel.
EquityDeals
A full-stack real-estate investment platform for presenting discounted property opportunities and supporting buyer engagement across a React and TypeScript interface with Firebase and backend services.