Project Info

Recall

Inspiration

At hackathons, conferences, and networking events, we meet dozens of people — yet often struggle to remember who they were, what we talked about, or even where we met them. We wanted to build a tool that helps you recall meaningful human interactions naturally — without needing to pull out your phone, take notes, or exchange LinkedIn profiles mid-conversation.

What it does

Recall is a wearable device that helps you capture and organize real-world social interactions in a privacy-respectful, effortless way. • When you say a natural voice cue like “Hi, I’m Ina!”, Recall automatically activates. • It uses a Raspberry Pi-based camera to detect the person you’re speaking with and generates a unique facial embedding (no images are stored — only anonymized vectors). • It then records and transcribes the conversation locally until the interaction naturally ends. • The transcript is sent to Gemini, which summarizes the discussion and extracts key details — names, affiliations, topics, and interests. • All this structured data — transcript summary, keywords, and face embedding — is securely stored in Supabase. • On your connected web dashboard, you can later query interactions conversationally: “Who was the girl who mentioned working at Amazon and loved pizza?” Recall then surfaces the right person and summary — so you can follow up or reconnect effortlessly.

How we built it

Hardware: Raspberry Pi 5 with ArduCam camera module and on-device microphone. • Triggering Logic: Voice activation detects natural conversation cues to start and stop recording. • AI Stack: • Face recognition: OpenCV + FaceNet for generating embeddings. • Speech-to-Text: Google Speech API. • LLM summarization & extraction: Gemini 1.5 Flash for concise structured summaries. • Backend: Supabase (PostgreSQL + Auth + Storage). • Frontend: Next.js web app for recall search and interaction visualization.

Challenges we ran into

Achieving low-latency face embedding generation on the Pi while maintaining real-time voice responsiveness. • Handling privacy responsibly — we designed Recall so it only activates when explicitly triggered, and only embeddings (not raw video) are stored. • Fine-tuning Gemini prompts to extract useful context (e.g., company names or shared interests) without losing nuance.

Accomplishments we're proud of

Built a fully functional wearable prototype within 36 hours that performs face recognition, transcription, and summarization in real time. • Created a natural, consent-based trigger system that respects privacy and minimizes passive surveillance. • Designed a clean, searchable dashboard that makes human memory queryable.

What we learned

Integrating multimodal AI (vision + audio + language) on embedded hardware is hard — but combining them thoughtfully can make technology feel human-centric again. We also learned that designing for consent and trust must come before cool tech.

What's next

for Re:call Add spatial mapping of interactions (time + location) for timeline visualization. • Improve on-device summarization for offline mode. • Build an API for CRM or personal knowledge base integration (Notion, Obsidian, etc.). • Explore smaller form factors (badge, glasses, or lapel pin).

Analysis

Compare with all teams

No indexed repository for this project, so there are no commit stats to show.

Technology

Found in codeNot checked

Google GeminiUnchecked
PythonUnchecked
SupabaseUnchecked

No repository was indexed for this project, so these Devpost claims have not been checked against code.

AI coding agents

No repository was indexed, so agent usage could not be checked.

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

No repository was indexed, so there is no codebase to measure.