Pratyush Ranjan Tiwari
@PratyushRT
Followers
1K
Following
7K
Media
79
Statuses
1K
Building privacy-preserving personal AI @eternisai, lots of RL lots of reward hacking, prev. PhD @JohnsHopkins, 3X EF cryptography grantee, built @ketlxyz
Joined November 2018
We introduce a better recipe for collecting post-training data when using GRPO. Collecting samples from experts is expensive, annotation budgets are limited. Which examples are actually worth paying for? We find that focusing on hard samples results in a 30-40% improvement. 1/7
3
51
329
RT @PratyushRT: Everything will be open-source for our models here. - Models released on huggingface already ✅.- Dataset live on huggingfa….
0
3
0
RT @PratyushRT: New blogpost: Reinforcement Learning for Privacy. We post-train small language models (SLMs) to be as good as GPT 4.1 at re….
0
12
0
Everything will be open-source for our models here. - Models released on huggingface already ✅.- Dataset live on huggingface already ✅.- Codebase for RL/GRPO with LLM Judge will soon be live 🚧. I'll answer any questions here or in DMs about using this model or learning RL.
1
3
11
RT @freysa_ai: 1/.Enchanted Mobile ✨. Your AI. On your phone. Private by default. Not one controlled by a big lab. Not one that hoards y….
0
21
0
We just launched Enchanted, a personal AI app that is private by default. Our goal is to bring Signal’s guarantees to personal AI. You can chat with the best models, and your queries are either run in TEE-GPUs or passed through the anonymizer.
0
1
8
Link to the full blogpost: Models coming to @huggingface in the next hour!. 8/8.
freysa.ai
Enabling sovereign AI and self-owned cognition at global scale.
1
0
6
The results:. - Qwen3-4B reached 9.55/10 anonymization score. - Nearly matching GPT-4.1 itself at 9.77/10. - Performance is practical: <250ms to first token, <2s total on consumer hardware. 7/8
1
0
7
Why not just use annotations? Because semantically-simialr replacement often has multiple correct answers. There isn’t one “gold” replacement for a name or number. This is why we pair GRPO with a live LLM judge (GPT-4.1), scoring each output in real time instead of relying only.
1
0
7
We built a dataset of ~30k examples with nuanced rules:. - Replace personal names with culturally similar ones.- Do not replace city names unless very small.- Never replace public figures like celebrities. This forces the model to learn subtle distinctions. 5/8.
1
0
7
Example: a whistleblower query about fraud at a big tech company gets transformed so the provider sees only a generic version. You still get legal advice, but the model provider cannot link the query back to you. 4/8
1
0
8
Our approach: surgical anonymization with semantic preservation. - Local model detects sensitive info.- Replace only what is private with context-preserving placeholders.- Send anonymized query through a TEE proxy.- Response is automatically deanonymized locally. 3/8
1
0
8
Why is this needed? While encrypted in transit, queries to model providers are still logged and can be leaked. The largest closed-source models are not offered in TEE-GPUs. Today, privacy often means choosing between capability or confidentiality. 2/8.
1
0
7
New blogpost: Reinforcement Learning for Privacy. We post-train small language models (SLMs) to be as good as GPT 4.1 at replacing sensitive info before queries ever leave your device. Goal: use closed-source LLMs without giving up on privacy. 1/8
7
12
37
RT @PratyushRT: We introduce a better recipe for collecting post-training data when using GRPO. Collecting samples from experts is expensiv….
0
51
0
We're working on many related problems as well as a broader mission at Eternis. One of our core products is a privacy-presercing personal AI interface. Most of the problems involve deeply understanding and improving LLMs, post-training models for cases where it makes the most.
0
1
3
RT @rohanpaul_ai: Training Group Relative Policy Optimization, GRPO, models on the hardest problems delivers the biggest gains when annotat….
0
5
0