DEV Community

Cover image for "TrailLog: use AI to spend less time using AI"
Padma Dhanush
Padma Dhanush

Posted on

"TrailLog: use AI to spend less time using AI"

#devchallenge #hf26challenge

Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass Submission 🌿

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

Most AI keeps us on our phones. TrailLog is built to do the opposite.

TrailLog is an offline field-guide agent that turns a walk into a mission. Tell it how much time you have ("I have 30 minutes") and it plans a short outdoor quest based on what you haven't spotted yet. Out on the trail, point your phone at a plant or tree, or record a bird call, and it identifies the species on-device and saves it to a private journal. When you're back, it asks for a quick reflection and uses it to shape your next quest.

The screen takes about 5 seconds per sighting. The walking, looking, and listening are the real experience.

Who it's for: beginner naturalists, hikers, run clubs, and students who want to learn local species without needing signal or staring at a phone.

Most apps measure success by how long you stay. TrailLog measures success by whether you leave.

Demo

Code

GitHub logo kanna857 / hh

FraudNexus AI

Closed-Loop Agentic Fraud Investigation & Next-Best Action Platform

FraudNexus AI is an enterprise-grade autonomous fraud investigation platform powered by TigerGraph relationship intelligence, agentic evidence seeking, GraphRAG historical case memory, deterministic compliance policy enforcement, and formal FinCEN SAR filings.

📑 Table of Contents

  1. Problem
  2. Solution
  3. Why Graph Intelligence
  4. Why Agentic Architecture
  5. Dataset & IEEE-CIS Profile
  6. System Architecture
  7. TigerGraph Graph Model
  8. GraphRAG & Case Memory
  9. Evidence Seeking Engine
  10. Policy & Approval Engine
  11. Next-Best Action Decisioning
  12. Benchmark Results
  13. Installation & Setup
  14. Environment Variables
  15. Running Locally
  16. Running Benchmarks
  17. Running Demo Lab
  18. Deployment Guide
  19. Limitations & Future Roadmap

🎯 Problem

Modern fraud operations teams suffer from alert fatigue and fragmented context. Traditional machine learning models output isolated risk scores without explanation, requiring human analysts to spend 30–60 minutes per alert manually searching database logs, querying device registries, matching cross-card relationships, and verifying regulatory policy compliance.

💡 Solution

FraudNexus AI automates the entire investigation…

View on GitHub

How I Built It

Tech stack

  • AI core: Gemma (open-weight), running locally via [Ollama / llama.cpp] for plant and tree identification and quest generation
  • Audio ID: BirdNET (open source)
  • Agent: Python multi-agent pipeline (Identifier, Journal, Quest Planner)
  • Memory: SQLite (local, private). Past sightings and reflections shape future quests.
  • Interface: [Streamlit / PWA]
  • Voice (optional): ElevenLabs delivers the quest as audio so the phone can stay in your pocket
  • Observability: Sentry Agent Tracing for latency, tokens, and tool calls per run
  • Development evidence: agent sessions shared via Entire / DevRelay (below)

How it works


text
"I have 30 minutes" + interests
  ↓
Quest Planner (Gemma + journal memory)
  ↓
Quest (text, or spoken via ElevenLabs)
  ↓
You close the app and go outside
  ↓
Photo or bird call → Identifier (Gemma / BirdNET, on-device)
  ↓
Journal (local SQLite)
  ↓
Short reflection → shapes the next quest

## Why Does Open Innovation Matter?
 **Works offline:** trails rarely have signal, and local inference means the app works anyway.
- **Private by default:** photos, recordings, location history, and reflections never leave the device.
- **Swappable models:** change the vision model or species set for your region without waiting on a vendor.
- **No cost to run:** no API key or per-call fees, so a school club or run group can use it freely.
- **Customizable agent:** quest logic is plain code and prompts you can change (difficulty, distance, season).

A closed API would have made the offline, private version impossible. The open model is what lets the app work in the place it's meant for: outside.

## My Agent Session

What the session shows:
- Choosing the idea and why it fits "Touch Grass"
- Picking the open stack (Gemma, BirdNET, SQLite)
- Building the Identifier, Journal, and Quest Planner
- [A bug or dead end you hit and how you fixed it]
- [What you changed after the field test]

## Prize Categories
- **Best Use of Gemma:** Gemma runs locally for on-device identification and quest generation.
- **Best Use of Entire:** agent sessions are shared above.
- **Best Use of ElevenLabs:** [spoken quests / demo narration]
- **Best Use of Sentry Agent Tracing:** [traces/screenshots of latency, tokens, tool calls]
<!-- Team Submissions: Please pick one member to publish the submission and credit teammates by listing their DEV usernames directly in the body of the post. -->

<!-- Thanks for participating! -->
Enter fullscreen mode Exit fullscreen mode

Top comments (0)