Production AI agents, and a faster CRM, for a VIP matchmaking company
AI agents that chat with clients, coordinate dates and book calls, plus the performance work that keeps the CRM behind them fast.
- Client
- Blush
- Through
- Qualkode Technologies
- Role
- Senior Software Engineer
- When
- Jan 2025 – present
Built as a Senior Software Engineer at Qualkode Technologies, for Blush.
- documents examined by the booking-availability query
- 2.18M → 1,684
- faster lookup in a function that kept timing out at 120 s
- 5.7×
- production-support tickets fixed at the root
- 45
The situation
Blush runs a white-glove matchmaking service. Matchmakers work out of a CRM, and clients talk to the service over iMessage, SMS and Instagram. AI agents carry much of that conversation: they answer, coordinate dates and book calls. That only works if they are right about dates, availability and who is talking, and if the CRM answers fast enough for the agents to rely on it.
What I built
- Built and run the production AI agents for profile chat and date coordination, with server-side guards that reject LLM-miscomputed dates and fake bookings, and Slack escalation to a human whenever a promise can't be kept.
- Rewrote the matchmaker-availability query: 2.18M documents examined down to 1,684 (about 1.76 s → 484 ms per call), and fixed a silent double-booking bug with transactional conflict checks.
- Rescued a production function that kept timing out at 120 s, where one lookup scanned 13.17M documents. The query fix alone made it 5.7× faster; with an index it is sub-second.
- Fixed a 77 MB response that was crashing a function on memory.
- Built AI-drafted social DM outreach with a human approval gate and centrally managed prompts.
- Built the CRM's unified iMessage and SMS inbox, calendar and booking, call recordings, and faceted profile search with typeahead.
Stack
- TypeScript
- React
- Serverless functions
- MongoDB
- OpenAI
- Vercel AI SDK
- Langfuse
- Slack API
