🚀 Inspiration
During my work, I noticed a recurring bottleneck: we are great at collecting feedback but terrible at processing it. I watched teams spend days manually categorizing Slack threads and spreadsheets. It was mentally draining and slow. I asked myself: “What if we could treat human feedback like structured data?” This led to Insightify, built to bridge the gap between "hearing" and "understanding" using the near-instant reasoning of the Gemini 3 Flash model.
💡 What it does
Insightify is an intelligent distillation layer that transforms messy, unstructured text into a high-level strategic roadmap.
- Thematic Clustering: Automatically groups recurring enterprise topics.
- Sentiment Pulse: A pulse check on the emotional tone of stakeholders.
- Actionable Roadmaps: Turns vague complaints into concrete 30-day tasks.
- Persistent History: A session-based archive to compare reports in real-time.
🛠️ How I built it
I optimized the project for speed and precision using the 2026 Google ecosystem:
- Engine: Integrated the Gemini 3 Flash-Preview model for ultra-low latency ().
- Frontend: Built with Streamlit 1.28.0+, featuring custom CSS3 injections for a "Command Center" aesthetic.
- The Logic: I implemented a multi-pass analysis that calculates a Priority Score () for every insight:
Where = Frequency, = Sentiment intensity (-1 to 1), and = Level of effort.
⚡ Challenges I ran into
- The "Ghost Box" UI: I had to debug deep CSS selectors to ensure the input headers stayed inside the white text area for visibility.
- API 404s: I initially struggled with model identifiers, eventually realizing the v1beta registry requires the specific
gemini-3-flash-previewstring. - Hallucination Prevention: To ensure accuracy, I forced the model to cite direct quotes as evidence for every generated roadmap item.
🏆 Accomplishments that I'm proud of
- Reliable Structured Output: Successfully forced the model to return valid, parsable JSON 100% of the time via system instructions.
- Stakeholder-Ready Exports: Implemented a one-click download system for professional plain-text reports.
- Speed: Achieving frontier-level reasoning at 3x the speed of previous generations.
📚 What I learned
- Vibe Coding: Realized that a great AI tool shouldn't feel like a cockpit; it should feel like a search bar for your own data.
- Token Optimization: Learned how to handle massive feedback volumes within the 1M+ token context window of Gemini 3.
- Secure DevOps: Mastered TOML secrets management to keep my API keys safe in a public repo.
Built With
- css3-custom-injection
- gemini-3-flash-preview
- github
- google-genai-sdk-2026
- python-3.11+
- streamlit-1.28.0+
- streamlit-community-cloud
- toml-secrets-management
