SIGIL

Home page
Signal Input
Graphical representation of outputs
Stock analysis and persona based summary
Use of Antigravity
Use of eleven labs

Images for Figma: https://drive.google.com/drive/folders/1Ej8W5zN0Q0fFHSwoCfJ7gutWe6hy2Y-E?usp=sharing

Inspiration

Financial insights are often hidden in unstructured text—news, market commentary, or everyday conversations—but turning those signals into actionable analysis is difficult. We built SIGIL to explore how AI models like Gemini can extract meaningful signals from raw text and combine them with real market data to generate clear, explainable investment perspectives inspired by investors like Warren Buffett and Ray Dalio.

What It Does

We leveraged Natural Language Processing (NLP) to convert raw market data text into a recommendation pipeline. The raw text first goes through Gemini’s API for signal extraction, allowing us to analyze:

Sector: which industry the information relates to (for example, tech, energy)
Direction: whether the signal suggests the stock may go up or down.
Timeframe: how long the signal may matter (i.e., short term vs long term.
Reliability: how trustworthy or plausible the information is.

It then enriches these signals with real financial data, filtering stocks within the S&P 500 so the analysis focuses on major publicly traded companies. By selecting the top 4 most relevant companies’ stocks, fetching historical price returns using the Massive API, and analyzing news sentiment from Alpha Vantage, it generates appropriate charts using MatPlotLib and a final Gemini-generated investment recommendation.

Persona-Based Recommendations: Users can also choose from personas (namely Ray Dalio, Warren Buffet, Simplify, and Quant) that change the tone and framework of the final recommendation. It is still the same data, but under different analytical lenses.

In-Line Chat Panel: After the results are loaded, an expandable Gemini chat panel appears below the analysis, pre-loaded with the summary as context, so that users can easily ask follow up questions about the specific results.

Signal Reliability Scoring: We added a reliability_score field to the signal extraction to flag obviously implausible inputs (e.g., "Google is going bankrupt" scores 5–15) versus highly actionable ones (e.g., overheard earnings beat → 85–95).

Accessibility: We also integrated browser-native Speech Recognition for voice input on the landing page, along with ElevenLabs Text-to-Speech for reading the final recommendation summary and Speech-to-Text in the Gemini chat assistant. This allows our website to be accessible to a wider range of end-users.

Challenges we ran into

API reliability and integration issues: We faced intermittent failures and inconsistent responses from external services like ElevenLabs, Massive, and Alpha Vantage. These included rate limits, latency, and response format differences, which made it difficult to reliably fetch market data, sentiment signals, and speech processing results in a single pipeline.

Handling noisy or ambiguous natural language inputs: Users can input vague or unrealistic claims, which made it challenging for Gemini to consistently extract accurate signals such as sector, direction, and timeframe. We addressed this by introducing a reliability scoring system to filter out implausible inputs before generating recommendations.

Real-time performance/latency: Because the system calls multiple APIs and performs analysis steps sequentially, keeping the total response time low while generating charts and recommendations was a challenge during the hackathon time frame.

Assumptions :

Raw text data contains actionable financial signals: We assumed that purely unstructured text - Market reports, overheard conversations, customer feedback which may or may not contain latent investment signals that an LLM can reliably extract with sector-level granularity.

We also assumed that mapping signals to one of the 11 standard GICS (Global Industry Classification Standard) would be specific enough for stock selection. We re scoped stock selection to S&P 500 constituents only, assuming sufficient sector coverage and liquidity.

Sentiment correlates with near term direction: We also took the Alpha Vantage news sentiment scores for selected stock that would complement return data in producing a useful recommendation.

Not all signals are trustworthy: We later took into consideration and implemented the fact that all inputs are NOT valid with the help of a reliability score letting the user know how reliable the output is.

Accomplishments (that we're proud of)

Full end-to-end pipeline: By splitting work effectively, we were able to connect NLP signal extraction with real market data, sentiment analysis, visualizing, and AI-generated recommendations and chat assistance in one seamless workflow within a short timeframe.

Designing persona-based financial analysis: Allowing users to interpret the same data through different investment philosophies (inspired by Ray Dalio and Warren Buffett) made the system both educational and engaging.

Designing for maximum accessibility: By leveraging ElevenLabs for including Text-to-Speech and Speech-to-Text, our website is able to improve its user experience by making it more interactive and accessible.

Reliability scoring for extracted signals: This helped prevent unrealistic or misleading inputs from producing misleading recommendations and added a layer of transparency to the system.

What we learned

How to leverage multiple AI and financial data services: This project required working with multiple APIs, including Gemini, Massive, and Alpha Vantage. Learning how to integrate all of these together into one seamless website was a steep learning curve, but this knowledge goes a long way to preparing each of us in working with more advanced APIs in the future.

The challenges of extracting reliable signals from natural language: Even powerful models like Gemini can struggle with ambiguous or speculative inputs, which reinforced the importance of validation mechanisms like reliability scoring.

Balancing technical complexity with usability: It is important to consider that technical complexity does not always equate to a good UX/usability. Therefore, we had to decide early on which technologies we needed, and which ones were stretch goals. Although we all were motivated to pursue complex technologies, ultimately we decided it was to the best of our end users’ interests to not over-complicate the tech stack.

What's next for SIGIL?

Expand beyond the S&P 500: Future versions could analyze mid-cap and international stocks, enabling broader market coverage.

Add portfolio-level recommendations: Instead of suggesting individual stocks, SIGIL could generate diversified portfolios aligned with the user’s risk tolerance and investment horizon.

Enhance real-time data analysis: Integrating live market feeds and faster caching mechanisms would reduce latency and make recommendations more responsive.