README.md

metadata

title: FutureSelves
emoji: ✨
colorFrom: yellow
colorTo: gray
sdk: gradio
sdk_version: 5.50.0
app_file: app.py
pinned: false
tags:
  - backyard-ai
  - openbmb
  - nvidia-nemotron
  - tiny-titan
  - best-agent
  - off-brand
  - bonus-quest-champion

✦ FutureSelves

A daily ritual where your future self sends you transmissions.

A private future-radio, not a dashboard. Check in with one word, receive a voice transmission from across time, make a tiny choice that reshapes who gets to speak tomorrow. The whole ritual lives inside a single signal chamber — no card-grid, no SaaS, no scroll.

All inference runs on-device via three small models — no cloud dependencies, no API bills, no data uploaded.

How it works

1. Onboarding — Tell the system about your current life chapter: what you're avoiding, what you're afraid won't happen, what's draining you, and what would make a miraculous year. Or skip straight to the demo with ✦ Try Maya's example — a fully-populated 4-day history from Maya, a 28-year-old founder, with audio for each transmission.
2. Daily check-in — One word + optional note for today. A structured insight extractor (Nemotron-Parse) reads your note for emotional signals.
3. Transmission — Your assigned future self (MiniCPM 2.5B, prompted with your full context) generates a personalized narrative message with a specific action prompt and cliffhanger. The transmission is spoken aloud via local TTS (Kokoro / Piper).
4. Your move — Choose: toward, steady, release, or repair. Each choice shifts your timeline and builds toward unlocking new cast members.
5. Reaction — Tell your future self how it landed. The next transmission remembers.

Models

Model	Params	Role	Sponsor
MiniCPM 2.5 (openbmb)	~2.5B	Transmission generation (primary LLM)	OpenBMB
Nemotron-Parse (NVIDIA)	<1B	Structured note extraction (emotions, themes, entities)	NVIDIA Nemotron
Piper / Kokoro	15–82M	Text-to-speech (pre-rendered samples + local synthesis)	—

Each model is well under 32B params. Total: ~3.1B across all three models — qualifies for Tiny Titan.

Prizes targeted (honest list)

Prize	Why we qualify
Backyard AI (track)	Practical daily-life app for personal reflection and emotional accountability. The demo persona (Maya) is a real-archetype user with 4 days of history, demonstrating depth.
OpenBMB MiniCPM Build	MiniCPM 2.5 is the primary generation model — not a side experiment. Every transmission in the live demo runs through it.
NVIDIA Nemotron	Nemotron-Parse extracts structured insights from check-in notes (sentiment, emotions, themes). Falls back to keyword extraction when GPU memory is tight — documented and graceful.
Tiny Titan	~3.1B total across all models — deep under the 4B cap.
Best Agent	Multi-step agentic pipeline: check-in → note extraction → prompt assembly → generation → choice → reaction → memory persistence. Open trace at traces/agent-trace.jsonl.
Off Brand	Custom transmission-console interface (Fraunces + IBM Plex Mono), atmospheric gradients and scanlines, animated tuning-state waves, voice-orb constellation rail, horizontal signal-path progress bar — a private future-radio, not a card-grid dashboard.
Best Use of Modal	Persona summarizer runs as a Modal serverless GPU function (traces/modal_app.py). Generates a 1-paragraph narrative summary from a persona's full history. Eligible for the $20k Modal credits pool.
Bonus Quest Champion	Six bonus criteria: 🔌 Off the Grid (no cloud APIs), 🎨 Off-Brand (custom UI), 📡 Sharing is Caring (open agent trace), 📓 Field Notes (blog post), + Nemotron sponsor + Modal sponsor.

Note on badges we don't claim: We do not claim 🎯 Well-Tuned (we use base MiniCPM 2.5, not a fine-tuned variant) or 🦙 Llama Champion (we use 🤗 Transformers, not llama.cpp). We also do not claim OpenAI Codex as a sponsor — the code was written by humans with AI assistance, not by Codex.

Tech

• UI: Gradio 5.50.0 with a custom transmission-console skin (Fraunces serif for the chamber voice, IBM Plex Mono for instrument labels), atmospheric gradients and scanlines, a horizontal voice-orb constellation rail, and a signal-path progress bar in place of step pills
• LLM: MiniCPM 2.5 via 🤗 Transformers with SDPA attention, running on a T4 GPU in HF Spaces
• Extraction: Nemotron-Parse (NVIDIA) with keyword fallback when GPU is constrained
• TTS: Piper 15–60M voices (pre-rendered samples in audio/voices/, served as static assets) + Kokoro 82M for live synthesis when available
• State: In-memory session state with gr.BrowserState persistence across page refreshes
• Demo persona: Maya, a 28-year-old founder with 4 days of pre-written transmission history. Pre-rendered audio for each. Click ✦ Try Maya's example on the first screen to skip onboarding.
• Agent trace: Every transmission logs the full chain (system prompt, user prompt, raw LLM output, parsed JSON, note insights, duration) to traces/agent-trace.jsonl for the Sharing is Caring bonus quest. See traces/agent-trace.jsonl.
• Modal integration: Persona summarizer runs on Modal's serverless T4 GPU. Function source in traces/modal_app.py, summaries in traces/persona-summaries.json.

Running locally

pip install -r requirements.txt
python app.py

Links

• Live Space — deployed on T4 GPU
• Source (monorepo)
• Demo video — 2-minute walkthrough
• Demo walkthrough — step-by-step guide for judges
• Social post — Medium article
• Agent trace — open record of every transmission
• Field Notes blog post — what we built and what we learned
• Modal app — serverless persona summarizer