PDAI Project Reference Guide

Approach · Dataset · Tech Stack · GitHub References

1. Fact Checking Support System

Approach Build a browser extension or web app where editors paste or submit content before
publishing.
Pipeline: claim extraction → evidence retrieval (web search + trusted sources) →
veracity scoring → report with flagged claims and source links.
Use LLM to decompose text into atomic claims, then retrieve evidence for each via
search APIs.
Return a confidence score per claim plus links to authoritative sources.

Dataset AVeriTeC (real-world claim verification dataset) — publicly available

LIAR dataset (12K labeled fake/real news statements) — publicly available
PolitiFact API / ClaimBuster API for real-time claim checking
Google Fact Check Tools API — free tier available
Tech Stack Python + LangChain / LangGraph for orchestration
OpenAI GPT-4o or Claude API for claim extraction and reasoning
Serper API / Google Search API for evidence retrieval
Streamlit or React frontend; FastAPI backend

GitHub [Link]/Libr-AI/OpenFactVerification — Loki: full pipeline (claim detection →

evidence → verdict)
[Link]/multimodal-ai-lab/DEFAME — multimodal fact-checker (text + images),
ICML 2025
[Link]/TheAtticusProject/cuad — legal contract NLP, useful for structured claim
analysis
[Link]/jagilley/fact-checker — simple LLM-based fact-checking with prompt
chaining

2. Talk to Your Finances

Approach A conversational interface over a user's transaction history (uploaded CSV or
connected via Plaid API).
NL queries get converted to SQL or pandas operations against the transaction table.
Features: spending summaries, category breakdowns, budget tracking, anomaly
alerts, natural language filters.
Optional: chart generation on demand ('show me a pie chart of last month's
spending').

Dataset User's own bank export (CSV/OFX) — no external dataset needed for prototype
Synthetic banking datasets: Kaggle 'Bank Transaction Dataset' — publicly available
Plaid API for live bank connection (sandbox mode available for testing)

Tech Stack Python + LangChain agents with SQL/pandas tool

SQLite or DuckDB for transaction storage
OpenAI function calling to route NL queries to data operations
Streamlit or React + Recharts for visualizations

GitHub [Link]/Kirushikesh/Personal-Finance-Agent — LLM agent with SQLite, NL-to-

SQL, AutoGen
[Link]/ShiweiHe0713/MoneyChat — GPT-powered chatbot connected to bank
statements with charts
[Link]/dantelmomsft/personal-finance-assistant-java — multi-agent banking
assistant (Java + Semantic Kernel)
 [Link]/RasaHQ/financial-demo — Rasa-based banking chatbot with pre-built
intents and flows

3. Work Security Inspector

Approach Mobile-first app for architects conducting site inspections.
Features: dynamic checklists (customizable per project type), photo capture with AI
hazard tagging, voice-to-text notes, auto-generated PDF inspection reports.
AI layer: upload a photo → computer vision detects PPE violations, structural hazards,
clutter.
Communication module: flag issues, assign to construction team members, track
resolution status.

Dataset OSHA construction safety dataset (public domain, US government)

PPE detection datasets on Roboflow Universe — free/open access
Construction Hazard Dataset (GitHub: roboflow community) — publicly available
No large proprietary dataset needed; can use YOLOv8 pretrained + fine-tune on small
set

Tech Stack React Native or Flutter for cross-platform mobile app

YOLOv8 (Ultralytics) for real-time PPE/hazard detection in photos
GPT-4o Vision API for open-ended hazard description
Firebase for real-time sync + photo storage; PDF generation with ReportLab

GitHub [Link]/topics/construction-safety — PPE detection with YOLOv8, helmet/vest

detectors
[Link]/ultralytics/ultralytics — YOLOv8 framework used widely for PPE detection
projects
SafetyCulture iAuditor (inspiration, not open source) — leading commercial reference
[Link]/topics/safety-monitoring — multiple AI-powered workplace safety
monitoring repos

4. Signability Checker
Approach User uploads or pastes a rental contract (PDF or text). System parses it and flags:
— Abusive or non-standard clauses (e.g., unusual repair obligations, excessive
penalties)
— Ambiguous or confusing legal jargon (explained in plain language)
— Missing standard tenant protections required by local law
Output: risk-scored clause-by-clause report with plain-language explanations and
recommendations.
Optional: jurisdiction selector to apply country/region-specific legal standards.

Dataset CUAD (Contract Understanding Atticus Dataset) — 510 contracts, 41 clause types,
publicly available at [Link]/TheAtticusProject/cuad
LEDGAR dataset — 60K+ contract provisions, Hugging Face Hub
Synthetic rental contracts (GPT-generated) for fine-tuning or few-shot prompting
Local tenant law documents (scraped from government sites — public domain)

Tech Stack Python + PyMuPDF / pdfplumber for PDF parsing

Claude or GPT-4 with structured prompts for clause classification and risk scoring
LegalBERT or DeBERTa fine-tuned on CUAD for clause extraction
Streamlit or [Link] frontend; color-coded risk report UI

GitHub [Link]/TheAtticusProject/cuad — gold standard for contract clause extraction

NLP
[Link]/LexPredict/lexpredict-lexnlp — legal NLP library (dates, clauses, named
 entities)
[Link]/OssamaLouati/Legal-AI_Project — full legal doc analysis web app ([Link]
+ Flask + CUAD)
[Link]/fanta-mnix/nlp-contract-analysis — NLP-based contract analysis in Python

5. AI Trends Explorer
Approach Automated pipeline that continuously ingests from: arXiv, Papers With Code, Hacker
News, major AI blogs, GitHub Trending.
LLM layer: summarize each item, extract key concepts/entities, cluster into themes.
Weekly digest: top emerging themes, notable model releases, benchmark
breakthroughs.
Search interface: users can query across the aggregated knowledge base using RAG.

Dataset arXiv API (free, no auth needed) — [Link], [Link], [Link] categories
Papers With Code API (free) — links papers to code repos and benchmarks
Hacker News API (free) — real-time tech discussions
RSS feeds from: OpenAI, Anthropic, DeepMind, Hugging Face blogs — all public

Tech Stack Python + feedparser / arxiv library for ingestion

LangChain + ChromaDB / Pinecone for RAG over aggregated content
OpenAI or Claude API for summarization and trend synthesis
Streamlit dashboard or Notion integration for digest output

GitHub [Link]/finaldie/auto-news — multi-source LLM news aggregator (RSS, Reddit,

YouTube, Tweets)
[Link]/hrnrxb/AI-News-Aggregator-Bot — Telegram bot aggregating arXiv,
Papers With Code, GitHub Trending
[Link]/AKAlSS/AI-News-Aggregator — Python RSS aggregator → Notion digest
[Link]/fabiochiusano/Awesome-AI-News — curated AI news with summaries and
sentiment

6. PowerEditor
Approach AI-powered writing assistant embedded in a web editor tailored to a team's house
style.
Features: tone/style consistency checker, auto-suggest rephrasings, template library,
grammar + editorial checks.
Style learning: ingest past approved content to build a custom style profile.
Collaboration: commenting, version history, role-based access (writer / editor /
approver).
Integration: export to PDF, email, or CMS directly.

Dataset Team's own past content (internal — no external dataset needed)

Common style guides (AP, Chicago) encoded as rules
Writing Prompts dataset (Reddit) for diverse writing samples if training needed —
publicly available
No large labeled dataset required; few-shot prompting with style examples is sufficient

Tech Stack React + TipTap or [Link] for rich text editor

Claude or GPT-4 API with custom system prompt encoding team style
[Link] / FastAPI backend; PostgreSQL for document storage
WebSockets for real-time collaboration (or Yjs CRDT library)

GitHub [Link]/nicolewhite/[Link] (use Tiptap instead) — [Link] is the leading

open-source rich text editor
[Link]/ueberdosis/tiptap — extensible rich text editor (React/Vue), widely used for
 AI writing tools
[Link]/steven-tey/novel — Notion-style AI-powered editor built with Tiptap +
OpenAI
[Link]/makeplane/plane — open-source project + docs management (reference
architecture)

7. AI Games
Approach Design a game where the core mechanic IS a machine learning concept — not just AI
characters.
Example concepts: 'descend the gradient' (navigate a loss landscape), 'train the
model' (data labeling puzzle), 'adversarial attack' (fool a classifier as a game
mechanic).
Alternative: adaptive narrative game where an LLM generates story branches based
on player choices, with no two playthroughs alike.
Keep the ML concept educational but fun — abstract it into game metaphors.

Dataset No external dataset required for most game concepts

For ML-concept games: generate synthetic data on the fly (e.g., gradient descent
visualized as terrain)
For NPC narrative: LLM API provides all dynamic content
MNIST or CIFAR-10 (public domain) if building a 'train the classifier' game mechanic

Tech Stack Unity or [Link] (browser-based) for game engine

OpenAI / Claude API for adaptive narrative generation
Python + matplotlib for ML concept visualizations (gradient descent terrain)
WebGL / [Link] for 3D loss landscape visualization games

GitHub [Link]/topics/ai-game — AI-driven game projects across multiple genres

[Link]/openai/gym — RL environments (reference for ML-mechanic game
design)
[Link]/AIDungeon/AIDungeon (inspiration) — original AI-narrative adventure
game
[Link]/topics/procedural-generation — generative content games using ML
models

8. Product Comparison
Approach User enters a product category or specific products to compare. System:
— Scrapes / retrieves specs, prices, and reviews from multiple sources
— Uses LLM to normalize specs across products into a unified comparison schema
— Generates a structured comparison table + narrative summary
— Allows user to tune weights ('I care most about battery life and price') to rerank
results.
Transparent sourcing: every claim links back to the original source.

Dataset Amazon Product Dataset (McAuley Labs) — publicly available, reviews + metadata
Open Food Facts / Open Product Data (for non-tech categories) — public domain
Scraped specs from manufacturer websites (legal for personal/research use)
SNAP Amazon Reviews dataset — Hugging Face Hub, publicly available

Tech Stack Python + Playwright / BeautifulSoup for web scraping

LangChain with structured output for spec normalization
Claude or GPT-4 for narrative comparison generation
React + Recharts for interactive comparison UI with adjustable sliders

GitHub [Link]/nicholasjclark/product-comparison-app — React product comparison UI

 reference
[Link]/mcauley-lab/AmazonReviews2023 — large-scale Amazon review +
metadata dataset
[Link]/topics/product-comparison — several open-source product comparison
tools
[Link]/apify/crawlee — robust web scraping library used for price/spec collection

9. Timetravel
Approach System that monitors published news articles and detects when claims become
outdated.
Two modes: (1) real-time check on new articles before publish, (2) retroactive scan of
archive.
Pipeline: extract time-sensitive claims → search for more recent evidence → flag
outdated statements → suggest updated sources.
UX: reader-facing highlights on outdated sentences with tooltip showing current
information.

Dataset CC-News (CommonCrawl News dataset) — public, multi-year archive of news articles
MIND (Microsoft News Dataset) — publicly available
Google News RSS / NewsAPI for current state of facts
Wikipedia revision history API (free) — tracks how facts change over time

Tech Stack Python + spaCy for temporal expression extraction (dates, 'currently', 'recently')
Bing News Search API or Serper for retrieving updated information
GPT-4 for outdatedness judgment and update suggestion
Browser extension (Chrome/Firefox) for reader-facing highlights

GitHub [Link]/Cartus/Automated-Fact-Checking-Resources — comprehensive fact-

checking resource list
[Link]/codelucas/newspaper — Python library for news article parsing and
extraction
[Link]/topics/temporal-information-extraction — NLP tools for time-sensitive claim
detection
[Link]/Libr-AI/OpenFactVerification — pipeline adaptable to temporal claim
verification

10. Taking Back Control of Social Media

Approach Browser extension that intercepts the social feed API responses and re-ranks them
according to user-defined preferences.
User writes natural language preferences: 'show me more AI research, less political
news, no ads, no rage-bait'.
LLM scores each post against user preferences, then reorders the feed in real time.
Filter controls: block keywords, boost topics, set content toxicity threshold.

Dataset No labeled training dataset required — uses LLM zero-shot scoring

Twitter/X Academic API or Reddit API for content (for testing pipeline)
HatEval / TweetEval datasets (Hugging Face) — toxicity/sentiment classifiers — public
User's own feed data (captured client-side) — no external dataset needed

Tech Stack Chrome Extension (Manifest V3) with content scripts to intercept feed
Claude or GPT-4 API for post scoring against user preferences
JavaScript + IndexedDB for client-side user preference storage
Optional: local ONNX model (DistilBERT) for offline scoring to reduce API calls

GitHub [Link]/topics/chrome-extension + social-media-filter — several feed filtering

 extensions
[Link]/huginn/huginn — agent-based news/social aggregation and filtering
platform
[Link]/ml-tooling/best-of-ml-python — reference for lightweight NLP classifiers for
filtering
[Link]/unitaryai/detoxify — open-source toxicity classifier usable for feed filtering

11. City Data Explorer

Approach Ingest municipal open datasets (e.g., Barcelona Open Data portal) and make them
accessible via:
— Interactive map visualizations (incidents, transport, demographics by neighborhood)
— Natural language query interface: 'What are the noisiest neighborhoods in
Barcelona?'
— Actionable citizen services: nearest facilities, service disruption alerts, event
listings.
Data pipeline: scheduled ingestion → normalization → vector store for NL queries +
geospatial DB.

Dataset Barcelona Open Data: [Link] — free, official, multiple

formats
European Data Portal ([Link]) — open datasets from EU municipalities
OpenStreetMap (Overpass API) — free geographic data
Eurostat urban statistics — free, public domain

Tech Stack Python + GeoPandas / Shapely for geospatial processing

PostGIS or DuckDB spatial extension for geospatial queries
Mapbox GL JS or [Link] for interactive maps
LangChain + RAG over dataset descriptions for NL query interface
Streamlit or [Link] dashboard

GitHub [Link]/cityofnewyork/open-data (reference) — mature open city data

infrastructure
[Link]/opendatasoft/ods-documentation — open data portal patterns and best
practices
[Link]/gboeing/osmnx — Python for street networks and urban data from
OpenStreetMap
[Link]/topics/urban-data — geospatial city data analysis projects

12. PressAnalyzer
Approach Tool that ingests articles or entire publications and scores them against established
press standards.
Scoring dimensions: source diversity, claim-to-evidence ratio, balance of perspectives,
factual accuracy signals, emotional language use, headline vs. body alignment.
Output: per-article scorecard + publication-level aggregate dashboard.
Can be used by media watchdogs, educators, or as a reader transparency tool.

Dataset MBIC (Media Bias/Fact Check) dataset — public, labeled media sources by bias
AllSides Media Bias Ratings — publicly available
SemEval 2019 Task 4 (hyperpartisan news detection) — publicly available
GossipCop / PolitiFact datasets (UPFD benchmark) — publicly available on GitHub

Tech Stack Python + Newspaper3k for article ingestion

Claude or GPT-4 for multi-dimensional rubric scoring via structured prompts
Hugging Face classifiers for bias/sentiment as supplementary signals
 Streamlit or React dashboard with radar charts per scoring dimension

GitHub [Link]/KaiDMML/FakeNewsNet — PolitiFact + GossipCop labeled news dataset

[Link]/crim-ca/media-bias-detection — media bias detection research codebase
[Link]/topics/media-bias — NLP bias detection projects
[Link]/codelucas/newspaper — Python article scraping and parsing library

13. Model Explainer

Approach System that wraps any ML model and produces dual-level explanations:
— Intuitive NL explanation: 'This loan was denied primarily because of high debt-to-
income ratio'
— Formal diagnostic: SHAP values, feature importances, partial dependence plots
with annotations
LLM translates the formal metrics into stakeholder-appropriate language.
Supports: tabular models (sklearn), neural networks, and provides counterfactual
explanations ('if X changed, the prediction would be Y').

Dataset UCI ML Repository (income, credit, medical datasets) — public domain

Kaggle competition datasets relevant to domain — mostly public
No additional training data needed; works on any user-provided model + dataset

Tech Stack Python + SHAP + LIME for model-agnostic explanations

Scikit-learn for model training; supports any sklearn-compatible model
GPT-4 / Claude for translating SHAP output into NL explanations
Streamlit for interactive UI with force plots and explanation narratives

GitHub [Link]/slundberg/shap — the SHAP library, gold standard for ML explainability

[Link]/marcotcr/lime — LIME: model-agnostic local explanations
[Link]/SeldonIO/alibi — comprehensive model explanation library
(counterfactuals, anchors)
[Link]/topics/explainability — XAI projects combining LLMs with formal metrics

14. Smart Wardrobe

Approach Connected wardrobe prototype with: (1) RFID or weight sensor layer for item tracking,
(2) mobile app for outfit planning, (3) AI styling assistant.
Features: outfit suggestion based on weather + calendar, wear frequency tracking,
sustainability insights (items never worn), packing list generator.
AR try-on component: overlay outfit combinations on user photo.
For prototype: simulate sensor data; focus on the app + AI layer.

Dataset DeepFashion dataset (CUHK) — 800K clothing images, attributes, landmarks —

publicly available
Polyvore Outfit dataset — outfit compatibility modeling — publicly available
OpenWeatherMap API (free tier) — weather-based outfit recommendations
User's own wardrobe data (photos + metadata) — no external dataset needed

Tech Stack React Native app + camera for wardrobe photo capture
GPT-4o Vision for clothing item recognition and tagging
Python + scikit-learn for outfit compatibility scoring (or fine-tuned model on Polyvore)
Firebase for wardrobe item storage; OpenWeatherMap API for context

GitHub [Link]/xthan/polyvore — Polyvore outfit dataset + compatibility model

[Link]/sidooms/DeepFashion — DeepFashion attribute recognition and retrieval
[Link]/topics/virtual-wardrobe — wardrobe app projects on GitHub
[Link]/Kjuly/iPokeMon (UI reference for mobile item inventory pattern)
15. Expat Assistant
Approach Personalized relocation assistant that generates dynamic checklists and advice based
on destination country, visa type, and personal situation.
Covers: visa/immigration steps, banking setup, healthcare registration, housing,
schools, cultural tips, tax obligations.
LLM layer: takes user profile (origin country, family size, profession) + destination →
generates prioritized, sequenced checklist.
RAG over up-to-date official government sources for accuracy.

Dataset Expatistan / Numbeo cost-of-living data — publicly available

Official government immigration portals (scraped, public domain)
InterNations community knowledge base (public articles)
Wikipedia country profiles + Wikidata — free and public

Tech Stack Python + LangChain with web search tool for up-to-date government info
RAG over scraped immigration guides using ChromaDB
Claude or GPT-4 for personalized checklist generation
React or Streamlit frontend with collapsible checklist UI

GitHub [Link]/topics/relocation — relocation tool projects

[Link]/public-apis/public-apis — lists cost-of-living, country info APIs for
integration
[Link]/nicholasgasior/expat-guide (reference) — community-driven expat advice
patterns
[Link]/Cinnamon/kotaemon — open RAG framework adaptable for document
Q&A on immigration guides

16. Market Research Assistant

Approach User inputs a research topic or domain. System:
— Retrieves recent papers (arXiv), news, reports, and competitor websites
— Extracts key trends, players, pain points, and market signals
— Synthesizes findings into a structured report: executive summary, trend map,
competitor matrix
— Allows follow-up queries: 'Who are the top 5 startups in this space?'
All claims are grounded with citations to source documents.

Dataset arXiv API, Semantic Scholar API — free access

Crunchbase Public Data (limited free tier) for startup intelligence
Google News / Bing News API for recent market signals
PitchBook / CB Insights reports (for reference; paywalled but summaries are public)

Tech Stack Python + LangChain agents with web search + arXiv tools
ChromaDB / Pinecone for RAG over retrieved documents
Claude or GPT-4 for report synthesis with citations
Streamlit or React for interactive report display with source drill-down

GitHub [Link]/assafelovic/gpt-researcher — autonomous LLM research agent (web +

papers → report)
[Link]/Stanford-OVAL/WikiChat — cited, grounded conversational research tool
[Link]/topics/market-research-ai — AI market research automation projects
[Link]/langchain-ai/langchain — LangChain framework used in most research
agent projects
17. Mindful Tourism
Approach Travel recommendation app that avoids 'tourist trap' suggestions and emphasizes
local, mindful, sustainable experiences.
User inputs destination + travel style + pace preferences → system generates itinerary
with local context.
Features: off-the-beaten-path suggestions, crowd timing advice, local etiquette tips,
accessibility info, eco-impact scores.
In-trip mode: contextual notifications based on GPS location + time of day.

Dataset OpenStreetMap + Overpass API — free, granular POI data worldwide

Yelp Open Dataset (limited countries) — publicly available
TripAdvisor public reviews (scraping or official partner API)
Atlas Obscura (public listings) — unusual and offbeat destinations

Tech Stack React Native for cross-platform mobile app with GPS
Mapbox GL JS for map rendering; Overpass API for POI retrieval
LangChain + Claude for itinerary generation and local context
Firebase for user preferences and trip storage; push notifications for in-trip mode

GitHub [Link]/topics/travel-app — open-source travel recommendation projects

[Link]/gboeing/osmnx — street-level urban data from OpenStreetMap
[Link]/nicktindall/cyclinguk-travel (reference) — sustainable travel planning
patterns
[Link]/topics/itinerary-planner — AI-powered trip planning projects

18. Real Estate Predictor Assessor

Approach Diagnostic dashboard for real estate ML models that surfaces where and why models
fail.
Features: geographic error maps (actual vs. predicted price heatmap by zone), error
distribution by property type/size/age, SHAP-based feature attribution per prediction.
Fairness analysis: check if model systematically under/over-values certain
neighborhoods.
Side-by-side model comparison: upload two models, compare their error profiles
spatially.

Dataset Kaggle House Prices dataset — public domain, classic benchmark

Zillow ZTRAX (academic license, free for research)
OpenDataSoft Real Estate transactions (France, Spain, etc.) — publicly available
Ames Housing Dataset (Iowa, 80 features) — public domain, ideal for feature
attribution demos

Tech Stack Python + scikit-learn / XGBoost for model training and prediction
SHAP for feature attribution; GeoPandas for spatial error analysis
Plotly + Mapbox for interactive choropleth error maps
Streamlit or Dash for the interactive diagnostic dashboard

GitHub [Link]/slundberg/shap — SHAP library for model explainability

[Link]/jorisvandenbossche/geopandas — geospatial analysis for mapping
prediction errors
[Link]/ResidentMario/missingno — data completeness visualization (useful for
feature analysis)
[Link]/topics/real-estate-prediction — real estate ML projects with various
explainability tools