// Built during my role as Senior AI/ML Engineer at Brainfog Technologies
IRIS Global
I built a full-stack, multi-agent AI research platform for NuqiWealth covering global stock markets across seven exchanges: NYSE, Nasdaq, NYSE American, London Stock Exchange, Cboe Europe, Dubai Financial Market, and Abu Dhabi Securities Market.
Tech Stack
How it works
The same seven specialized agents behind IRIS India, fundamentals, technicals, sentiment, ranking, screening, broad search, and portfolio, handle each type of question, routed through a layer that decides which agents a given question actually needs.
- Built seven specialized agents: fundamental, technical, sentiment, ranking, screener, broad search, and portfolio, each owning a different slice of the analysis.
- Built an automatic answer-quality check that scores every response and falls back to a live web search when internal data isn't sufficient.
- Built a morning scheduler that automatically analyzes hundreds of stocks across all seven exchanges: running cross-agent analysis, storing results in the database, running sector-level analysis and a health checklist per sector, and backtesting the technical agent's own decisions against the last 5 years of data.
- Connected the platform directly to a user's real portfolio and watchlist for personalized analysis.
Results
Specialized Agents
Global Exchanges Covered
Faster After Backend Optimization
Before
A comprehensive cross-agent query on a single stock took up to 2 minutes. Every specialized agent wrote out a full analysis independently, an orchestrator combined those into one essay, and a formatting guardrail rewrote it again for brand consistency.
After
A 4-phase backend optimization (shared with IRIS India, on a separate codebase) collapsed that into a single high-quality synthesis pass at the end of the pipeline. The same comprehensive analysis now takes 49-55 seconds (about 55% faster), single-agent queries like fundamentals or technicals answer in 20-25 seconds, and general questions get a dedicated fast-path that skips the agent pipeline entirely, answering in 3-8 seconds. Sentiment analysis, previously the slowest single-agent query at up to 90 seconds, now answers in 35-45 seconds. Averaged across all query types, response time fell from roughly a minute to under 30 seconds.
Why I built it
Investors tracking global markets across multiple exchanges usually need several disconnected tools, and most AI research assistants either run every agent on every question (slow) or fail silently when their data is incomplete.
How I approached it
I built a routing layer that decides which specialized agents a question actually needs instead of running all of them every time, added a quality check that falls back to live web search when internal data falls short, and pre-computed overnight analysis so the platform feels instant during market hours.
Full System Architecture
The actual, code-verified architecture: the Docker/infrastructure layer, both nightly scheduler jobs, the main orchestrator graph, all six sub-agents, cross-agent synthesis, the MCP integration, and the portfolio/watchlist integration. Explore each part below, drag to pan and scroll or pinch to zoom.
