Lunatechs Meetup #11 · May 14, 2026
A packed room for the practical version of AI trading.
The room was full before the talks started: packed couches, people along the glass, a livestream camera, and the usual food-and-badges cluster.
I was one of the four organizers and hosts for the night. I was not just watching the talks. I was helping create the Hong Kong tech room I want more of: technical, curious, and willing to get into the details.
”AI trading” gets vague fast. This night did not. Four speakers opened laptops and showed workflows: SEC filings, research dashboards, signal testing, and natural-language trading agents.
Why this one was different
Less slideware, more working systems.
There was a clear disclaimer at the top. This was not financial advice. It was a technical workshop about using AI with financial data, research workflows, and software tools.
The best parts were about workflow design: deterministic code where exactness matters, LLMs where ambiguity is real, and guardrails before an agent becomes more than a demo.
Four workflows
Each talk showed a different place to put the agent.
Jason Xu
Python + LLMs for the junior-banker grind: fetch filings, map line items, reconcile statements, and generate a 3-statement model.
Sandeep Muthangi
AI agents as research infrastructure: market data, outlier reports, thematic shifts, stored notes, and visualizations.
Shally Liu
Claude Code for crypto factor research: data exploration, sentiment features, factor testing, and a cautious first backtest.
Dale Satre
Minara as an agentic trading interface: research chat, strategy studio, backtesting, paper trading, and defined execution rules.
The constraint
Use the LLM where ambiguity is real.
Jason gave the cleanest architecture lesson of the night: use deterministic code when the job needs one exact answer, and use the LLM where ambiguity is real.
A general agent building a model from a 10-K can waste time finding its footing. The better version was specific: Python fetches filings, code builds the structure, checks verify it, and the LLM helps with messy label matching.
The room
Builders, quants, founders, finance people, and curious first-timers.
Compass Offices hosted us on the 16th floor at Lee Garden Two in Causeway Bay. Lunatechs and Berkeley Club Hong Kong brought the crowd.
Questions went straight to the hard parts: temperature, restatements, signal quality, Claude Code versus Python scripts, and the line between research assistant and trading system.
After the talks
The pizza counter was the second room.
People stayed after the livestream ended, which is usually the real test. If everyone leaves when the last slide disappears, it was content. If people stay, it was community.
This had the better version: old friends, strangers comparing notes, repo requests, tool questions, and trading ideas explained over pizza.
Closing note
Hong Kong needs more rooms where demos can break in public.
AI in finance is easy to sell from a slide. It is more useful when someone opens the laptop, shows the script, and names the guardrails.
