Stocker
An AI-powered, Bloomberg-style stock analysis platform backed by 21 agents. Three live portfolios tracked against the S&P 500, 167 stocks to double-digit gains, 13 past 75%.
type
AI Product @ Finance
Role
Product & AI Engineer
platform(s)
Web
website
stocker.cx
Stocker was an AI-powered stock analysis platform I worked on between late 2025 and early 2026. It used a network of agents to analyze the market, research opportunities, and pick stocks, running against real market data every trading day.
Under the hood it was a 21-agent system working in two tiers. Twelve junior agents each covered a different corner of the market, scanning for opportunities, doing the research, and scoring the best candidates. Their output was the daily watchlist. Three senior analysts then picked from that watchlist, each running three agents internally to build positions, refine them, and write full research reports on every stock.
I was the sole engineer on the project, responsible for the multi-agent pipeline behind the picks, the interface that presented them, and the infrastructure that kept it all running.
The Watchlist
Stocker began with the watchlist, a system of AI agents that scanned the market for opportunities. It was where the project started, and where almost all of the gains came from.
The watchlist was where Stocker began. Twelve junior agents each covered a different corner of the biotech market, scanning daily for catalysts: FDA decisions, clinical trial results, insider activity, partnership announcements. Each one researched what it found, scored it, and submitted its results. What came out was a ranked shortlist of the strongest opportunities across the market that day.
This was the part that worked. The watchlist ran every trading day, building a record of which names the system was most confident in and why. Tracking those picks over time produced Stocker’s most striking results, and it became the foundation everything else was built on.
The Analysts
Once the watchlist was working, we extended the platform into three senior analysts. Each one was a multi-agent system in its own right, with its own investing personality. One was disciplined and diversified, one took high-conviction bets, and one chased momentum and short-term catalysts.
Each analyst picked from the watchlist, explained its reasoning, and ran its own portfolio. They were tracked in real time against each other and against the S&P 500, so you could watch three different strategies play out on the same set of opportunities. This was the newer layer of the platform. The watchlist was the proven engine, and the analysts were where we pushed it next.
What It Showed
The numbers came from the watchlist, and they were hard to argue with.
Over its run, the watchlist tracked 167 stocks to double-digit peak gains. 13 of them climbed more than 75%, including one that ran +290% in 29 days, another +210%, and another +187%. Most picks hit their peak within three to four weeks of being surfaced. That confirmed the system was finding real opportunities, not noise.
The Takeaway
Getting LLMs to produce consistent structured output at every step of the pipeline was its own problem. The normalization layer exists because raw agent submissions didn’t always conform to the schema, and one malformed result could break the entire downstream flow. Every handoff between agents needed validation, and that meant thinking carefully about failure modes before they happened in production.
The front end looked clean but the data behind it was anything but. Live prices came in over WebSocket, portfolio NAV was being recalculated every few minutes by a separate service, and the UI had to merge senior analyst reasoning with junior catalyst data, two different scoring sources, and a real-time decisions feed all at once. Most of the front-end work was in making that feel seamless.
The AI side and the product side had to work in lockstep. That coordination was most of the work. Whatever you’re trying to build, this is the level I work at.
