cd /news/artificial-intelligence/show-hn-time-to-text-time-series-ai-… · home topics artificial-intelligence article
[ARTICLE · art-114281] src=github.com ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

Show HN: Time-to-text time-series AI to limitate LLM hallucinations

An open-source hybrid local framework, RobinHood, combines TCN, LSTM, and Attention deep learning with Captum Integrated Gradients feature attribution to generate deterministic numerical weights for time-series data, which are then sent to open-source LLMs on Hugging Face to produce financial text reports, aiming to eliminate hallucinations in automated financial reporting. The framework runs 100% on-device for raw data processing and training, sending only anonymized weights to the cloud for text synthesis, and includes a sample gold price dataset for immediate testing.

read2 min views3 publishedAug 28, 2026
Show HN: Time-to-text time-series AI to limitate LLM hallucinations
Image: Michielbdejong (auto-discovered)

An open-source, hybrid local framework designed to analyze multi-asset time-series data. It leverages a deep learning architecture (TCN + LSTM + Attention) combined with a mathematical feature attribution layer (Captum Integrated Gradients) to eliminate hallucinations in automated financial reports.

📊

Quick Start:A ready-to-usegld_price_data.csv

dataset is included in the repository so you can test the framework instantly.

Traditional Large Language Models (LLMs) tend to be baffling—they are great at reasoning but poor at raw math. When processing raw numerical tables or dense time-series, autoregressive models often invent trends or hallucinate metrics, breaking structural analytical trust.

By reversing the usual approach and feeding the language interface with deterministic prediction attributes, we can force the model to strictly explain each of the AI's choices.

This framework isolates raw mathematical feature attributions on-device, delivering a cold, objective, and unbiased look at multi-asset data across custom horizons without computational guesswork or statistical fabrication.

Local Compute (100% On-Device): Your raw time-series data, deep learning training loops (TCN + LSTM), and Captum mathematical feature attributions run entirely locally on your hardware.Text Synthesis (Hybrid Cloud): Only the final, anonymized deterministic numerical weights are sent via API to open-source LLMs on Hugging Face to generate the human-readable text report.Your raw data never leaves your machine.

Follow these precise steps to deploy and execute the pipeline locally on your machine:

gh repo clone ikaruz050309/RobinHood
cd RobinHood

Ensure you have Python installed, then run the installation command to fetch all background packages:

pip install -r requirements.txt

To enable natural language synthesis from your local evaluation weights via Hugging Face Serverless Inference, export your API token:

export HUGGINGFACE_API_KEY="your_huggingface_api_key_here"
python3 main.py

If you want to upload your own custom data, your .csv

dataset must match the following technical parameters to prevent matrix dimensions or compliance checks from halting the pipeline:

CSV Format Only: The system exclusively parses structured.csv

files. - Single Company Constraint: The architecture evaluates standalone data matrices. Do not include a 'Ticker' column or mix data from multiple corporations, or the engine will immediately halt the execution loop. - Purely Numerical Time-Series: Except for a single chronologically sequential date column, all tracking columns must contain pure continuous float or integer sequences. - Data Density (The Richer, The Better): High data density directly translates to robust mathematical attribution maps. Ensure your historical record counts vastly exceed your configured context windows ($N$ ) and prediction horizons ($K$ ).

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @robinhood 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/show-hn-time-to-text…] indexed:0 read:2min 2026-08-28 ·