---
title: README
emoji: ๐ฐ๏ธ
colorFrom: gray
colorTo: blue
sdk: static
pinned: false
---
**PxA Labs.** An independent research collective from India building open, measurable intelligent systems, from safety monitors for fine-tuned LLMs and accent-robust speech recognition to interpretable genomics and real-time adaptive optics. Every result is reported against a fair baseline, including what did not work.
### Research areas
- **AI Safety & Alignment.** Monitoring and repairing safety drift in fine-tuned models.
- **Language & Speech.** ASR that closes the accent gap for Indian-accented English.
- **Bioinformatics.** Interpretable ML for regulatory signals in non-coding DNA.
- **Multi-Agent Systems.** Agent teams with observability and runtime validation built in.
- **Signals, Optics & Systems.** Adaptive optics, acoustic physical layers, fault-tolerant infrastructure.
### Models, datasets and demos
| Project | On the Hub |
|:--|:--|
| **SafetyPulse**: catches safety drift during fine-tuning, repairs it, audits the repair | Models: [tiny-aya-global-safetypulse-repaired](https://huggingface.co/purvansh01/tiny-aya-global-safetypulse-repaired) ยท [NF4 4-bit](https://huggingface.co/purvansh01/tiny-aya-global-safetypulse-repaired-nf4) ยท LoRA adapters ([alpaca](https://huggingface.co/purvansh01/safetypulse-lora-alpaca), [alpacacur](https://huggingface.co/purvansh01/safetypulse-lora-alpacacur), [gsm8k](https://huggingface.co/purvansh01/safetypulse-lora-gsm8k), [none](https://huggingface.co/purvansh01/safetypulse-lora-none))
Datasets: [eval](https://huggingface.co/datasets/purvansh01/safetypulse-eval) ยท [finetune-data](https://huggingface.co/datasets/purvansh01/safetypulse-finetune-data) ยท [pairs](https://huggingface.co/datasets/purvansh01/safetypulse-pairs) ยท [probe-bundle](https://huggingface.co/datasets/purvansh01/safetypulse-probe-bundle) ยท [logs](https://huggingface.co/datasets/purvansh01/safetypulse-logs) |
| **IndiVoice-DeepASR**: Whisper + LoRA for Indian-accented speech | Model: [whisper-indian-lora](https://huggingface.co/PxALabs/whisper-indian-lora) |
| **AutoMaintainer**: an always-on autonomous AI software engineering team | Space: [AutoMaintainer](https://huggingface.co/spaces/purvansh01/AutoMaintainer) |
Code for every project is on [GitHub](https://github.com/PxA-Labs).
### Connect
[Website](https://pxa-labs.pxalabs.workers.dev) ยท [GitHub](https://github.com/PxA-Labs) ยท [LinkedIn](https://www.linkedin.com/company/pxa-labs/) ยท [Instagram](https://www.instagram.com/pxa_labs) ยท [Discord](https://discord.gg/W4BQ8835tk) ยท pxalabs@gmail.com