Inspiration
The inspiration came from the classic "Turing Test" but flipped on its head. Instead of asking if a machine can think, we wanted to ask: "Can a machine lie well enough to fool a human?" We were fascinated by the psychological tension of real-life interrogations where a single logical inconsistency can break a case. We wanted to build a platform that turns "detective work" into a rigorous exercise for a user's logical thinking and decision-making skills.
What it does
Connect & Press: Interrogator is an AI-powered mystery simulator where users enter a high-stakes interrogation room to face suspects based on real-life case references. The suspect is powered by an LLM that tracks its own internal "Stress Level." Users must : Analyze case files and evidence (CCTV, GPS logs, forensics).Identify logical contradictions in the suspect's live testimony.Make split-second "Decision Moment" choices to press for a confession or offer a plea deal.Earn points based on the precision of their logic and the efficiency of their interrogation.
How we built it
We utilized a modern full-stack approach focused on real-time AI interaction: Frontend: Built with Next.js and Tailwind CSS for a gritty, immersive UI.AI Engine: Powered by Gemini 3 Flash, using advanced system instructions to manage the suspect’s persona and internal state.Database: Neon (PostgreSQL) stores case data, evidence logs, and user progress.Logic: We implemented a Stress Meter $S$ where $S \in [0, 100]$. The suspect's behavior transitions from calm to contradictory as $S$ increases based on valid evidence presentation.
Challenges
we ran into One of the biggest hurdles was ensuring the AI didn't "hallucinate" an early confession. We had to fine-tune the system prompt to ensure the suspect remained defensive and only broke when a specific logical contradiction was presented. We also faced challenges in managing API credit limits during intensive testing of the "Decision Moment" mechanics.
Accomplishments
that we're proud of We are incredibly proud of the "Thought Signature" feature. It gives users a "detective's intuition" by showing a glimpse into the AI's raw logical processing—allowing players to see the "Ghost in the Machine" as it tries to calculate its next lie. Achieving a seamless JSON-structured communication between the LLM and the UI for real-time stress updates was also a major win.
What we learned
Building this project taught us a lot about Prompt Engineering for stateful characters. We learned how to use LLMs not just for text generation, but as a game engine capable of tracking complex variables. We also deepened our understanding of UX for AI, specifically how to make an "invisible" stress meter feel impactful through dialogue changes.## What's next for Connect & Press: Interrogator We plan to introduce Multi-Suspect Interrogations, where users must cross-reference statements from two different AIs to find inconsistencies between them. We are also looking into integrating voice-to-text so users can physically speak their interrogation to the suspect, adding another layer of realism to the "Decision Moments."
Log in or sign up for Devpost to join the conversation.