A newer version of the Gradio SDK is available: 6.22.0
Build Small Hackathon — Master Strategy Report
For: Sushruth | Goal: Win prize money, cover max sponsors, build 3 apps
Hackathon: Build Small (Hugging Face × Gradio) | Status: LIVE NOW (June 2026)
1. Hackathon Overview
| Detail | Info |
|---|---|
| Host | Hugging Face × Gradio |
| Total Prize Pool | $48,000+ cash + physical prizes |
| Tracks | Backyard AI + Adventure in Thousand Token Wood |
| Format | Gradio Spaces on HF org build-small-hackathon |
| Model constraint | Small models ≤ 32B parameters |
| Participant credits | $250 Modal + $100 Codex + $20 HF — for every registrant |
| Awards stack? | YES — one app can win track prize + multiple sponsor prizes |
| Submissions | Submit as many apps as you like |
2. Complete Sponsor & Prize Map
| Sponsor | Prize | How to Qualify | Models/Tools |
|---|---|---|---|
| HuggingFace (Track) | 1st: $4k · 2nd: $2.5k · 3rd: $1.5k · 4th: $1k (per track) | Top placement in Backyard AI or Thousand Token Wood | Any small model ≤32B |
| OpenBMB | $5k pool per track ($10k total) | Use MiniCPM models as core brain | MiniCPM-V 4.6, MiniCPM-o 4.5, MiniCPM5-1B, MiniCPM4.1-8B, VoxCPM2 |
| NVIDIA | Two RTX 5080 GPUs (physical, ~$1,000 each) | Best space + community engagement using Nemotron | Nemotron-Nano, Nemotron-Omni, Nemotron-ASR, Nemotron-Parse, Nemotron-Embed |
| OpenAI | ChatGPT Pro subscription (1 year) | Codex-attributed commits in GitHub repo or Space history | OpenAI Codex (coding AI) |
| Modal | $20k credits for winners + $250 per participant | Build/deploy on Modal compute | Modal serverless GPU compute |
| Cohere | Part of prize pool | Use Tiny Aya or Cohere Transcribe | Tiny Aya (multilingual LLM) · Cohere Transcribe (ASR/voice) |
| Black Forest Labs | Prize (via track judge picks) | Use FLUX.2-klein for image generation | FLUX.2-klein (4B/9B diffusion) |
| Community Choice | $2,000 | Most likes/interactions on HF Space | Any — visual, shareable, emotional wins votes |
Bonus Merit Badges (bonus on top — collect them all)
- Tiny Titan: All models ≤ 4B
- Custom UI: Frontend beyond default Gradio (use
gr.Server) - Fine-tuned model: Publish a fine-tuned model on HF
- llama.cpp runtime: Run model via llama.cpp
- Agent trace: Share agent trace on Hub
- Blog post: Publish a write-up on HF blog
3. Your 3 Ideas — Evaluated
Your Idea 1: Smart Budget + Junk Food Tracker
Receipt photo → OCR → categorize spending + junk food → budget advisor
| Dimension | Score | Notes |
|---|---|---|
| Track fit | ⭐⭐⭐⭐⭐ | Perfect Backyard AI — real problem for a real person |
| Uniqueness | ⭐⭐⭐⭐ | No existing app combines budget + junk food in one small-model Gradio Space |
| Sponsor coverage | ⭐⭐⭐⭐⭐ | OpenBMB (MiniCPM-V) + Cohere (multilingual) + Modal (background) |
| Your skills fit | ⭐⭐⭐⭐⭐ | Your CV/OCR/anomaly detection + RAG experience is directly applicable |
| Buildability (3 days) | ⭐⭐⭐⭐ | Moderate — receipt OCR + budget logic + Gradio UI |
| Prize potential | ⭐⭐⭐⭐⭐ | Track prize ($4k) + OpenBMB ($2.5k) + Modal ($10k credits) + Cohere |
| Overall | 9.5/10 | Your strongest idea — build this first |
Your Idea 2: Kid Photo → Sketch + Painting Tool
Photo → FLUX.2-klein line art → in-app canvas painting for kids
| Dimension | Score | Notes |
|---|---|---|
| Track fit | ⭐⭐⭐⭐⭐ | Perfect Thousand Token Wood — whimsical, visual, kid-focused |
| Uniqueness | ⭐⭐⭐ | Many apps do photo-to-sketch; your angle is the in-app painting on HF |
| Sponsor coverage | ⭐⭐⭐⭐ | Black Forest Labs (FLUX.2-klein) + OpenBMB (optional MiniCPM tips) |
| Your skills fit | ⭐⭐⭐⭐ | Good — uses your CV/vision model experience |
| Buildability (3 days) | ⭐⭐⭐ | Canvas JS overlay in Gradio is tricky but doable |
| Prize potential | ⭐⭐⭐⭐ | Track prize + Black Forest Labs + Community Choice if shareable |
| Overall | 7.5/10 | Good secondary project for Thousand Token Wood |
Your Idea 3: Voice-Cloned Bedtime Stories
Mom records voice → MiniCPM generates story → VoxCPM2/XTTS narrates it in mom’s voice
| Dimension | Score | Notes |
|---|---|---|
| Track fit | ⭐⭐⭐⭐⭐ | Thousand Token Wood — emotional, whimsical, clearly AI-powered |
| Uniqueness | ⭐⭐⭐ | Many existing consumer apps (MomSays, Totora, MamaTales) do this, but none as small-model HF Space |
| Sponsor coverage | ⭐⭐⭐⭐⭐ | OpenBMB (MiniCPM + VoxCPM2) + Cohere (multilingual + Transcribe) |
| Your skills fit | ⭐⭐⭐⭐ | Good — uses your NLP + audio pipeline skills |
| Buildability (3 days) | ⭐⭐⭐ | Voice cloning setup is time-consuming; VoxCPM2 is easiest path |
| Prize potential | ⭐⭐⭐⭐⭐ | Track prize + OpenBMB + Cohere + Community Choice (parents share viral clips) |
| Overall | 8.5/10 | Strong secondary for Wood, especially with VoxCPM2 for OpenBMB alignment |
4. Top 15 Ideas — Master Strategy Table
Legend for Sponsor Coverage
MB = MiniCPM/OpenBMB · NV = NVIDIA Nemotron · CO = Cohere · BF = Black Forest Labs · MD = Modal · CX = OpenAI Codex
Track A: Backyard AI (Real problems for real people)
| Rank | App Name | Real Problem | Core AI Stack | Sponsors | Win Chance | Difficulty | Unique Angle |
|---|---|---|---|---|---|---|---|
| 1 | Smart Receipt + Budget Advisor (Your Idea 1) | Students/immigrants overspend on junk food; can’t manage monthly budget | MiniCPM-V (OCR + classify) + Cohere Tiny Aya (multilingual advice) + Modal (background inference) | MB · CO · MD | ★★★★★ | Medium | Combines budget + junk food + nutrition goals in one app. You = real user. Receipt photo is fast proof. |
| 2 | German Document Decoder | Immigrants/students can’t understand Behördenbriefe, rental contracts, insurance docs | MiniCPM-V (doc reading + summarize) + Cohere Tiny Aya (translate to Turkish/Arabic/Hindi) + VoxCPM2 (read aloud) | MB · CO | ★★★★★ | Easy-Med | Very relevant in Germany. You can demo with a real Aachen neighbor. Multilingual = Cohere prize sweet spot. |
| 3 | Lab Anomaly Detection Dashboard (Perplexity Idea 3) | RWTH lab or student workshop has sensor/image data but no AI anomaly detection | MiniCPM-V (visual inspection) + Nemotron-Nano (reasoning/explanation) + Modal (batch inference) | MB · NV · MD | ★★★★ | Hard | Directly uses your robotics/CV/anomaly detection skills. NVIDIA GPU prize if Nemotron is core brain. |
| 4 | Research Paper RAG Assistant (Perplexity Idea 1 enhanced) | MSc/PhD students overwhelmed by papers, need summaries, Q&A, study plans | MiniCPM (RAG + Q&A on PDFs) + Nemotron-Embed (embedding) + Cohere Transcribe (voice Q&A) + Modal | MB · NV · CO · MD · CX | ★★★★ | Medium | Uses your RAG thesis experience directly. Stack 4+ sponsors. Build with Codex → Codex prize. |
| 5 | Car Dashboard Warning Interpreter | Non-technical drivers terrified by warning lights on dashboard | MiniCPM-V (photo → identify lights + explain + urgency) + Cohere (multilingual) | MB · CO | ★★★★ | Easy | Near-zero competition on HF. Every family member with a car is a real user. Fast to build. |
| 6 | Local Shop Inventory AI | Small Aachen shop owner takes shelf photos → needs stock count + reorder advice | MiniCPM-V (count items from image) + Nemotron-Nano (business reasoning) + Modal (background) | MB · NV · MD | ★★★ | Medium | Targets NVIDIA Nemotron (GPU prize) + OpenBMB. Good B2B Backyard AI story. |
| 7 | On-Device Medication Helper (Elderly) | Elderly parents confused by medication names, side effects, doctor notes | MiniCPM-V (read prescription/label) + Cohere Transcribe (voice input) + VoxCPM2 (audio output) | MB · CO | ★★★★ | Medium | Highly emotional story. “I built this for my grandparent.” Full voice-in/out pipeline = Cohere showcase. |
| 8 | ADHD Focus Journal + Task Breaker | People with ADHD struggle with task management, brain dumps, daily planning | Nemotron-Nano (break tasks into micro-steps) + Cohere Transcribe (voice dump) + MiniCPM (weekly insights) | MB · NV · CO | ★★★ | Medium | Note: competition exists (NeuroBait is already in hackathon). Differentiate with Nemotron reasoning. |
Track B: Thousand Token Wood (Creative, whimsical, AI-as-the-core)
| Rank | App Name | Concept | Core AI Stack | Sponsors | Win Chance | Difficulty | Unique Angle |
|---|---|---|---|---|---|---|---|
| 9 | Dream Visualizer + Interpreter | User types or speaks a dream → AI interprets symbols → generates surreal art → reads it back | MiniCPM or Nemotron-Nano (symbolic analysis) + FLUX.2-klein (surreal art) + VoxCPM2 (narration) + Cohere Transcribe (voice input) | MB · NV · BF · CO | ★★★★★ | Medium | Covers 4 sponsors in one app. Dreams are deeply personal and shareable. NVIDIA GPU prize via Nemotron. Strong Community Choice candidate. |
| 10 | Voice-Cloned Bedtime Stories (Your Idea 3 enhanced) | Mom records voice → kid picks genre/theme → MiniCPM writes story → VoxCPM2 narrates in mom’s voice | MiniCPM5-1B (story gen) + VoxCPM2 (voice TTS with clone) + Cohere Tiny Aya (multilingual) | MB · CO | ★★★★★ | Med-Hard | VoxCPM2 is OpenBMB’s own TTS → perfect OpenBMB alignment. Emotional = viral = Community Choice votes. |
| 11 | Kid Photo → Sketch + Paint (Your Idea 2 enhanced) | Photo → line art sketch → kid paints on it in-browser | FLUX.2-klein + line art LoRA + MiniCPM (fun story about drawing) + custom JS canvas | MB · BF | ★★★★ | Med | Black Forest Labs prize alignment. Add story narration (MiniCPM) to differentiate from plain photo-to-sketch apps. |
| 12 | Interactive Branching Story + AI Illustrations | User picks genre + characters → Nemotron writes branching plot → FLUX.2-klein illustrates each scene | Nemotron-Nano (story + branching logic) + FLUX.2-klein (scene images) + Cohere Tiny Aya (multilingual) | NV · BF · CO | ★★★★ | Medium | NVIDIA GPU prize potential (Nemotron as story brain). Illustrated interactive fiction is visually stunning for demos. |
| 13 | AI Plant Doctor + Fantasy Portrait | Photo of plant → MiniCPM-V diagnoses health issue → FLUX.2-klein generates fantasy portrait of plant → care tips in plant’s “own voice” | MiniCPM-V (diagnosis) + FLUX.2-klein (portrait) + VoxCPM2 (plant voice TTS) | MB · BF | ★★★★ | Easy-Med | Extremely shareable (plant lovers post everywhere). Easy to demo. Covers Black Forest Labs + OpenBMB. Fast to build. |
| 14 | Micro-Fiction Zine Generator | 1-sentence prompt → 3-panel story (MiniCPM) → 3 illustrations (FLUX.2-klein) → downloadable mini-zine PDF | MiniCPM5-1B (story panels) + FLUX.2-klein (panel art) + Cohere (multilingual) | MB · BF · CO | ★★★ | Easy-Med | PDF download is practical. Artists and writers will share their zines. Good Community Choice candidate. |
| 15 | AI Neighborhood Simulator | Tiny MiniCPM agents (baker, doctor, café owner) live in a neighborhood; user actions trigger agent reactions | MiniCPM5-1B × multiple agents + Modal (background simulation loop) + Nemotron-Embed (agent memory) | MB · NV · MD | ★★★ | Hard | Very creative — if it runs smoothly, it’s unforgettable. But risk of overengineering. Only attempt if time allows. |
5. Recommended 3-App Strategy — Maximum Prize Money
The core principle: each app should stack 3+ prize categories. Awards stack completely.
App 1 — PRIMARY (Backyard AI) ⭐ Highest cash potential
Build: Smart Receipt + Budget + Junk Food Advisor
| Element | Detail |
|---|---|
| Track | Backyard AI |
| Real user | You or a student friend in Aachen — use real receipts from REWE/Aldi |
| Core model | MiniCPM-V 4.6 (OCR + item extract + junk food classify + advice) |
| Sponsor 2 | Cohere Tiny Aya — translate advice to user’s native language (Turkish/Hindi/etc.) |
| Sponsor 3 | Modal — run inference jobs in background (batch receipt processing) |
| Build with | OpenAI Codex — commit Codex-attributed code → qualifies for ChatGPT Pro prize |
| Badges to target | Tiny Titan (all models ≤4B) · Custom UI · Blog post |
| Demo proof | 5 real receipt photos → show junk food table, daily budget limit, monthly forecast |
| Prize stack | Backyard AI 1st ($4,000) + OpenBMB Backyard ($2,500) + Modal winner credits ($10,000) + Cohere pool + ChatGPT Pro |
| Estimated potential | ~$17,000+ in cash + credits |
Feature scope (3 days):
- Receipt photo upload → MiniCPM-V extracts store, items, prices as JSON
- Junk food classifier → table showing (chips, soda, cake) vs (bread, vegetables, etc.)
- Budget tracker → set €500/month → shows daily limit, remaining budget, overspend warning
- AI advice → “You spent €38 on chips this month. Cut to once a week → save €25/month.”
- Optional: Cohere Tiny Aya → all advice output in user’s native language
- Optional: Modal → run MiniCPM inference in background for batch receipts
App 2 — SECONDARY (Thousand Token Wood) ⭐ Emotional + multi-sponsor
Build: Dream Visualizer + Interpreter
| Element | Detail |
|---|---|
| Track | Thousand Token Wood |
| Concept | User types or speaks a dream → AI analyzes symbols → generates surreal art → narrates the interpretation |
| Core model | Nemotron-Nano (deep symbolic/psychological reasoning about dream content) |
| Sponsor 2 | FLUX.2-klein (Black Forest Labs) — surreal/dreamlike image of the dream |
| Sponsor 3 | VoxCPM2 (OpenBMB) — narrates the interpretation in a calm, atmospheric voice |
| Sponsor 4 | Cohere Transcribe — user speaks their dream instead of typing |
| Sponsor 5 | Modal — run FLUX generation as background job |
| NVIDIA angle | Nemotron-Nano as the reasoning core → qualifies for RTX 5080 GPU prize |
| Demo proof | Record yourself speaking a dream → show generated art + interpretation → clip for social media |
| Prize stack | Thousand Token Wood 1st ($4,000) + OpenBMB Wood ($2,500) + NVIDIA RTX 5080 GPU ($1,000) + Black Forest Labs + Cohere + Community Choice ($2,000) |
| Estimated potential | ~$10,000+ in cash + GPU + credits |
Feature scope (3 days):
- Text input OR voice input (Cohere Transcribe) of a dream
- Nemotron-Nano: analyze dream symbols → produce structured interpretation (themes, emotions, meaning)
- FLUX.2-klein: generate a surreal dream-scene image from key symbols
- VoxCPM2: narrate the interpretation in a calm, atmospheric voice
- Gallery: show past dreams + their art
- Share button → downloads art + text for social sharing (viral = Community Choice votes)
App 3 — BOOSTER (Thousand Token Wood or Backyard AI) ⭐ Cover remaining sponsors
Build: Voice-Cloned Bedtime Stories (Your Idea 3 — upgraded)
| Element | Detail |
|---|---|
| Track | Thousand Token Wood |
| Concept | Mom records voice (30 sec) → kid picks genre → MiniCPM generates story → VoxCPM2 narrates in mom’s cloned voice |
| Core model | MiniCPM5-1B (story generation — short, age-appropriate, genre-based) |
| Sponsor 2 | VoxCPM2 (OpenBMB’s own TTS/voice clone) — deepens OpenBMB prize case |
| Sponsor 3 | Cohere Tiny Aya — multilingual (generate + narrate story in German/Hindi/Arabic) |
| Why this as App 3 | Covers OpenBMB (VoxCPM2) and Cohere again from a different angle, and is highly viral for Community Choice |
| Demo proof | Record mom/friend’s voice → show kid UI → play story in their voice → parents will share the clip |
| Prize stack | Thousand Token Wood 2nd/3rd ($1,500–$2,500) + Community Choice ($2,000) + Cohere pool |
| Estimated potential | ~$4,000–$6,000 in cash + credits |
Feature scope (3 days):
- Voice recording UI (30-60 sec) — record or upload WAV
- VoxCPM2 voice clone creation from reference audio
- Kid-friendly genre selector: King, Animal, Space, Adventure, Fairy Tale
- MiniCPM5-1B: generate 2-minute age-appropriate story from genre + kid’s name
- VoxCPM2: TTS narration in cloned voice → audio player
- Optional: Cohere Tiny Aya → generate story in German/Hindi/Turkish
6. Prize Stacking Calculator (Realistic Estimate)
| Prize Category | App 1 (Budget) | App 2 (Dreams) | App 3 (Stories) | Total |
|---|---|---|---|---|
| Track 1st place | $4,000 | — | — | $4,000 |
| Track 2nd/3rd | — | $2,500–$4,000 | $1,500–$2,500 | $4,000–$6,500 |
| OpenBMB (per track) | $2,500 | $2,500 | — | $5,000 |
| NVIDIA RTX 5080 GPU | — | $1,000 (physical) | — | $1,000 |
| Modal credits (winner) | $10,000 | $7,000 | — | $17,000 |
| Cohere pool | Partial | Partial | Partial | ~$1,500 |
| Community Choice | — | $2,000 | $2,000 | $2,000–$4,000 |
| ChatGPT Pro (Codex) | ✓ | — | — | Subscription |
| Realistic total | $30,000–$38,000 in cash + credits + GPU |
Note: Modal credits are cloud compute credits, not cash. Still very valuable for your thesis work.
7. Design Guide — How to Make Each App Look Beautiful
Judges see hundreds of default Gradio apps. A great UI can win the Custom UI merit badge and push you from 3rd to 1st.
General Rules for All 3 Apps
- Use
gr.Serverwith custom CSS — move beyond default Gradio theme - Custom font (Google Fonts CDN via
gr.HTML) — Poppins or Inter looks clean - Color scheme: pick 2 accent colors + white/dark background
- Add app logo (generate with FLUX.2-klein itself for irony points)
- Mobile-friendly (HF Spaces run on mobile too — judges may check on phone)
App 1 — Budget Tracker
Color scheme: Deep navy (#1B2A4A) + Fresh green (#27AE60) + White
Layout:
- Top: Upload receipt button + camera icon
- Middle: 3-column dashboard (Spent / Remaining / Daily Limit) as big bold numbers
- Bottom: Junk Food Table (red items) vs Healthy Table (green items)
- Right panel: AI Advice chat bubble (conversational, not a wall of text)
- Make the budget numbers BIG (like a fitness app shows calories)
- Red/green color coding: junk = red row, healthy = green row
- Progress bar for monthly budget (like a fuel gauge)
App 2 — Dream Visualizer
Color scheme: Deep purple (#1A0533) + Electric blue (#4A90E2) + Gold (#F5C518)
Layout:
- Hero: Dark gradient background with floating particle effect (CSS)
- Center: Mic button (voice input) OR text area — minimal, atmospheric
- Result: Full-width dream art image with interpretation text overlaid at bottom
- Footer: Gallery of past dreams (polaroid card style)
- This app should LOOK like a dream — dark, mystical, premium
- Generate the app header art itself with FLUX.2-klein
- Play a soft ambient sound loop on load (optional, via gr.Audio autoplay)
App 3 — Bedtime Stories
Color scheme: Soft cream (#FFF8E7) + Sky blue (#5DADE2) + Moon yellow (#F9E04B)
Layout:
- Step 1: Big "Record Mom's Voice" button with moon/star icons
- Step 2: Genre cards with illustrations (4-5 colorful cards with kid icons)
- Step 3: Story player — large play button, star animation while playing
- Background: Subtle starfield animation (CSS keyframes)
- Use big rounded buttons, large font sizes (for kids but also for demo wow factor)
- Generate genre card illustrations with FLUX.2-klein
- Keep instructions in 3 steps only — no walls of text
8. Execution Plan (3 Apps in ~7 Days)
| Day | Tasks |
|---|---|
| Day 1 | Set up HF Space, Gradio scaffold, test MiniCPM-V API locally. Get first receipt OCR working for App 1. |
| Day 2 | Complete App 1 core (OCR + junk food classifier + budget logic). Add Cohere multilingual. Polish UI. |
| Day 3 | Deploy App 1 on Modal + HF Space. Record demo with real receipts. Publish blog post. |
| Day 4 | Build App 2 (Dream Visualizer) — Nemotron reasoning + FLUX.2-klein image gen. Voice input via Cohere Transcribe. |
| Day 5 | App 2 VoxCPM2 narration + UI polish (the dark atmospheric design). Add gallery. Deploy + share on social. |
| Day 6 | Build App 3 (Bedtime Stories) — VoxCPM2 voice clone + MiniCPM story gen + kid UI. |
| Day 7 | Final polish on all 3 apps. Write READMEs with frontmatter tags. Record demo videos. Post on social (Twitter/LinkedIn). |
9. Winning Checklist for Each App
Before submitting each Space, verify:
- Deployed in
build-small-hackathonorg on HF - README includes frontmatter tags (track, sponsors, model names)
- Demo video recorded and linked in README
- Real user proof in README (quote, screenshot, use case)
- GitHub repo connected with Codex-attributed commits (App 1 especially)
- Model size confirmed ≤ 32B (ideally ≤ 4B for Tiny Titan badge)
- Blog post published on HF blog (counts as merit badge)
- Social post shared (HF, Twitter/X, LinkedIn) to drive Community Choice votes
- Modal deployment configured (for App 1 and App 2)
- Cohere Transcribe or Tiny Aya integrated (App 1 + App 2 + App 3)
10. Quick Model Reference
| Model | Size | Use Case | Sponsor | API/Hosted |
|---|---|---|---|---|
| MiniCPM-V 4.6 | ~8B | OCR, image understanding, receipt reading | OpenBMB | HF hosted / OpenBMB API |
| MiniCPM5-1B | 1B | Text generation, story writing, advice | OpenBMB | Local / HF Spaces ZeroGPU |
| MiniCPM4.1-8B | 8B | Reasoning, complex instructions | OpenBMB | HF hosted |
| VoxCPM2 | Small | Voice cloning + TTS | OpenBMB | HF Space |
| Nemotron-Nano | 8B | Reasoning, deep analysis (dreams, anomalies) | NVIDIA | NVIDIA API |
| Nemotron-Embed | 340M | Embeddings for RAG/memory | NVIDIA | NVIDIA API |
| FLUX.2-klein | 4B/9B | Image generation, dream art, portraits | Black Forest Labs | HF Spaces ZeroGPU |
| Cohere Tiny Aya | 3.35B | Multilingual text, translation | Cohere | Cohere API (free tier) |
| Cohere Transcribe | — | Speech-to-text, voice input | Cohere | Cohere API |
| XTTS-v2 | 335M | Alternative voice cloning (if VoxCPM2 too slow) | Open source | HF: coqui/XTTS-v2 |
Report compiled June 2026. Hackathon is LIVE — start building today.