Get in touch
Work location
B-810/811, World Trade Tower (WTT), Makarba, Sarkhej Gandhinagar Highway, Ahmedabad,
Gujarat, India, 380015
Social link
Athena replaces hours of manual portfolio research with a single, structured recommendation generated by a coordinated team of specialized AI agents, rather than one generic chatbot.
We partnered with the team to bring that system to life end to end, from pipeline design through the streaming interface investors interact with. Every portfolio runs through a disciplined research-to-report pipeline that gathers live market data, runs technical and fundamental analysis, and commits to a graded call with a confidence score, target price and risk level.
Investors are flooded with fragmented signals: price charts, earnings filings, breaking news and technical indicators, scattered across a dozen tools. Generic AI chat tools can summarize a stock on request, but they don’t reason like an analyst, weighing conflicting signals and committing to a graded recommendation with a stated confidence level.
Athena replaces that fragmented process with a multi-agent analytical system. Each stage of real analyst work (research, analysis and reporting) is handled by a dedicated agent in a coordinated pipeline, producing a structured, evidence-backed portfolio report instead of a freeform chat answer.
Instead of asking an LLM to “analyze my stocks” and hoping for a coherent answer, Athena runs every portfolio through a disciplined research-to-report pipeline. It gathers live market data, runs technical and fundamental analysis, and compiles everything into a clear, actionable investment report with confidence scores and risk levels. Portfolios are tracked over time rather than analyzed once, so recommendations are revisited as holdings and market conditions change. It’s built for retail investors and advisors facing too much market noise and too little time to turn it into a decision.
Raw market signals move through four coordinated stages, each handled by a purpose-built agent with a narrow, well-defined job. This separation is what keeps the system explainable: at any point you can see which agent produced which piece of the final recommendation, and why.
Queries a Qdrant vector store of SEC filings, earnings reports and prospectuses on demand, so an agent pulls the exact passage it needs instead of relying on the model’s memory.
.svg)
Continuously pulls live market data, company fundamentals and relevant news, grounding every analysis in current, verifiable information.
Runs parallel technical and fundamental analysis: RSI, MACD and Bollinger Bands alongside DCF valuation modeling.
Synthesizes every signal into one structured output: a buy, hold or sell call, a confidence score, a target price and a risk level.
Agents run as nodes in a deterministic graph. Each stage passes structured state to the next, keeping the pipeline predictable, debuggable and easy to extend.
When a question needs source-grounded detail, an agent queries the Qdrant store and cites the retrieved passage directly rather than paraphrasing from memory.
Holdings are monitored over time. As prices move or new filings and news land, Athena revisits its recommendations and confidence scores.
The interface streams each agent’s reasoning as it happens, so users watch the system gather data, apply models and build its call live.
Agent state is checkpointed at every step, so the workflow can pause, recover or resume mid-analysis without losing context.
Agents, tools, orchestration and API layers are cleanly separated, so new data sources, models or agents can be added without touching the rest.
The aim was a decision a leadership team could defend, not a list of possibilities. Here is how the outcome differs from where most teams start.
The first question is never “where could AI go?” but “what are you about to decide, and what would make it defensible?”
The people running a process describe the real one, including the workarounds, which is usually where the opportunity is.
Collapsing them into one number hides the trade-off. High impact with low feasibility is a different conversation from the reverse.
Ideas we recommend against are listed with the reason, so the same suggestion doesn’t return in six months with nothing learned.
Value alone surfaced work the client had no capacity to start. Readiness became a scoring dimension, not a footnote.
Priorities, effort and tool recommendations are written so any competent team can pick them up. The build is a separate decision.
.jpg)
