Update README.md

README.md

@@ -13,9 +13,9 @@ thumbnail: >-
   <img src="https://cdn-uploads.huggingface.co/production/uploads/6a8fe39e88002faf4779f515/DAArlmahepwXthIoWa85J.jpeg" alt="Seerror" width="500">
 </p>
 
-<h3 align="center">The AI That Sees What You Can't.</h3>
-<p align="center"><b>India's First AI Privacy Suite — Free for Everyone</b></p>
-<p align="center"><i>Audit. Secure. Optimize. Protect.</i></p>
+<h3 align="center">Built for humans. Against the machine.</h3>
+<p align="center"><b>An Indian AI research company building small, safe AI that runs on-device</b></p>
+<p align="center"><i>India's Digital RAW</i></p>
 
 <p align="center">
   <a href="https://www.seerror.com">🌐 Website</a> ·
@@ -31,19 +31,38 @@ thumbnail: >-
 
 ## 🛡️ Who We Are
 
-Seerror is **India's Digital RAW**, an AI-powered privacy suite built in Bengaluru to protect 1.4 billion Indians from threats they don't even know exist: website vulnerabilities, surveillance, data leaks, scams, and AI dependence.
+Seerror is an Indian AI research company building small, safe AI models that run on-device, starting with security: protecting people from scams and fraud.
 
-**Built for humans. Against the machine.** No subscriptions. No compromise. Zero data collected. Zero investor money.
+**Built for humans. Against the machine.** The machine is everything that exploits people: scams, surveillance, data harvesting. We build AI that stands with humans against it.
 
-> *We don't file reports. We neutralise threats.*
+- 13 products live, free to use
+- On-device AI that collects no user data
+- Our own models, datasets and research papers
+- Fully bootstrapped, no investor money
 
 ## 🎯 The Mission
 
-Every day, people across India lose money and trust to scams that could have been stopped in seconds, if the right information had reached them in time. Seerror builds AI that protects people from digital fraud: offline, on-device, and without compromising their privacy.
+Every day, people across India lose money and trust to scams that could have been stopped in seconds, if the right information had reached them in time. Our models run on your own phone or computer, so protection does not depend on a connection and your questions never leave your device.
+
+We publish our failures as well as our results, because trust in AI has to be earned.
 
 ## 🤖 Our Models
 
-- **[Seerie](https://huggingface.co/seerror-technologies/seerie-GGUF)** — an offline AI cybersecurity assistant trained on real Indian cybercrime scenarios. Understands Hindi, English, Hinglish, and multiple regional Indian languages. Runs entirely on-device with zero internet required. [Learn more →](https://seerror.com/seerie-model.html)
+**Seerie** is an offline AI assistant that helps people in India recognise scams and know what to do next, in English, Hinglish and Indian languages. It is a fine-tuned Qwen3-1.7B, released as GGUF files that run on-device.
+
+| Model | Trained on | Paper |
+|---|---|---|
+| [seerie-v349-GGUF](https://huggingface.co/seerror-technologies/seerie-v349-GGUF) (latest) | dataset v349, 108,405 conversations | *Three Times the Data, the Same Mistakes* |
+| [seerie-v224-GGUF](https://huggingface.co/seerror-technologies/seerie-v224-GGUF) | dataset v224, 36,669 conversations | first evaluation paper |
+
+Both papers report a hand-checked audit next to the automatic scores, including the answers the model got wrong. [About Seerie →](https://seerror.com/seerie-model.html)
+
+**In progress:** the Seerie training corpus has grown past 140,000 conversations, now including native Hindi (Devanagari). Every law-section citation and every unverified phone number or website has been removed from the training data, and the next model is evaluated on a hand-written test set that the training data has never seen.
+
+## 🔬 Research
+
+- **Seerie papers:** preprints with evaluation data and audits on Zenodo, DOI [10.5281/zenodo.23110075](https://doi.org/10.5281/zenodo.23110075) (all versions).
+- **GARUDA:** on-device AI for surveillance drones. It sees, thinks and sends honest alerts, while a person always makes the decision. It has no weapon function. GARUDA is in development and tested on simulated data; it has not flown yet.
 
 ## 🧰 The Seerror Suite
 
@@ -65,19 +84,19 @@ Every day, people across India lose money and trust to scams that could have bee
 | Tool | What it does |
 |---|---|
 | [Seercode](https://seerror.com/extension/seercode.html) | Catches Aadhaar, PAN, UPI and other Indian PII, plus API keys, before you commit |
-| [Seerpack](https://seerror.com/extension/seerpack.html) | Scores npm packages 0–100 for typosquatting, malicious scripts and supply-chain risk |
+| [Seerpack](https://seerror.com/extension/seerpack.html) | Scores npm packages 0 to 100 for typosquatting, malicious scripts and supply-chain risk |
 
 **Web**
 
-- **[Seerror AI](https://www.seerror.com)** — AI-powered website security audits. *(v2 in progress)*
+- **[Seerror AI](https://www.seerror.com)**: AI-powered website security audits *(v2 in progress)*
 
-All tools are **free forever**, run **on your device**, and collect **zero personal data**.
+All tools are **free**, run **on your device** where they can, and collect **zero personal data**.
 
 ## 🤝 Partners
 
 Proud member of the **[Sarvam Startup Program](https://seerror.com/sarvam.html)**.
 
-## 📚 Research & Resources
+## 📚 Resources
 
 [Research papers](https://seerror.com/research.html) · [Datasets](https://seerror.com/datasets.html) · [Docs](https://seerror.com/docs.html) · [Blog](https://seerror.com/blog/) · [Downloads](https://seerror.com/downloads.html)
 
@@ -85,4 +104,4 @@ Proud member of the **[Sarvam Startup Program](https://seerror.com/sarvam.html)*
 
 📧 [seerrorworks@gmail.com](mailto:seerrorworks@gmail.com) · 🔒 Found a vulnerability? Email us with the subject *"Responsible Disclosure"*.
 
-<p align="center"><sub>Founded by Jay Tiwari · Bengaluru, India · Since 2024</sub></p>
+<p align="center"><sub>Founded by Jay Tiwari · Bengaluru, India · Since 2024</sub></p>