Systems shipped under real load
Two production deployments — documented architecture, measured outcomes, no abstracted case studies. Scope and constraints are on the table.
E-commerce AI Chatbot
Designed for zero-hallucination output at 100K concurrent users. The retrieval pipeline runs on a Postgres-backed vector index — no third-party black boxes that drift under spike traffic.
Query optimization cut median response latency by 60 ms at peak load. Architecture decisions are documented: every constraint, every trade-off, no vague 'AI-powered' abstraction.
Stack: Python · FastAPI · PostgreSQL pgvector · AWS ECS · Redis
Scale target: 100K concurrent sessions · p99 latency under 120 ms
Real-Time Fintech Analytics
Built for a fintech client whose previous dashboard was refreshing on 15-minute batch cycles. The new system streams events through a Kafka pipeline — data reflects market state within 800 ms.
Database schema redesigned around time-series access patterns. Aggregation queries that took 4 seconds now resolve in under 200 ms at production row counts.
Stack: Node.js · Kafka · TimescaleDB · React · AWS MSK
Latency: event-to-display under 800 ms · query p99 under 200 ms
From the teams who ran these systems
The chatbot held through our Black Friday spike — 80K sessions in two hours, no degradation. The query architecture Jomart's team designed is the reason it didn't melt.
We'd lived with batch-refresh dashboards for two years. Switching to a live pipeline felt like finally seeing what was actually happening in our data — not a 15-minute-old approximation.
— Head of Engineering, mid-market e-commerce platform
— CTO, Series A fintech startup
Have a system that needs to hold at production scale? Let's talk architecture before we talk timeline.