Agent view
Hello, agent.
Everything on this site in plain text. Fetch it directly at /llms.txt, or use the RSS feed and sitemap.
# PxA Labs > PxA Labs is a multi-disciplinary research collective building high-fidelity intelligent systems across speech, AI safety, bioinformatics, multi-agent systems and signal processing. Contact: pxalabs@gmail.com · GitHub: https://github.com/PxA-Labs · Discord: https://discord.gg/W4BQ8835tk ## Founders - Purvansh Joshi (Co-founder, Dehradun, India): Full-Stack Developer | ML & Cloud Enthusiast | Research Scholar | Open Source Contributor. https://www.linkedin.com/in/purvansh-joshi/ · https://github.com/purvanshjoshi · https://portfolio-gray-one-24.vercel.app/ · https://huggingface.co/purvansh01 · https://x.com/PurvanshJoshi - Archit Mittal (Co-founder, Uttarakhand, India): Full-Stack Developer | Cloud & Open-Source | Building Scalable Tech for Social Impact. https://www.linkedin.com/in/archittmittal/ · https://github.com/archittmittal · https://www.architmittal.me · https://x.com/archittmittal ## Research domains - AI Safety & Alignment: Monitoring and enforcing safe behaviour in models and agents — from refusal circuits to deterministic policy engines. - Language & Speech Intelligence: Auditory interfaces and ASR that close regional accent gaps for Indian-accented speech. - Bioinformatics: Interpretable machine learning that decodes regulatory signals in non-coding DNA on reproducible, modest compute. - Multi-Agent Systems: Agent teams that act inside real tooling — with observability and runtime validation built in, not bolted on. - Signals, Optics & Systems: Real-time adaptive optics, acoustic physical layers and fault-tolerant data infrastructure. ## Projects - [SafetyPulse](https://pxa-labs.pxalabs.workers.dev/projects/safetypulse) (AI Safety, Active): SafetyPulse checks whether a model's internal refusal mechanism is still intact during fine-tuning, using a Circuit Integrity Score (CIS) computed from 50 prompt pairs. It gates each checkpoint PASS/WARN/FAIL, repairs failing models with SafeTune, and verifies the repair with a paired audit (AuditKIT). Reported results — ASR rise from benign Alpaca tuning: 14.3% → 25.2% on 119 HarmBench test prompts; CIS vs ASR correlation: Spearman ρ = −0.80 [−0.85, −0.71]. Code: https://github.com/PxA-Labs/safetypulse - [Project RIPRA](https://pxa-labs.pxalabs.workers.dev/projects/project-ripra) (Optics, Active): RIPRA reconstructs atmospheric-turbulence-distorted wavefronts from Shack-Hartmann sensor frames and computes deformable-mirror actuator commands for adaptive optics. It combines a C99 real-time core with ML reconstructors and an LSTM predictor for loop-lag compensation. Built for the ISRO Bharatiya Antariksh Hackathon 2026. Reported results — Hot-path latency per frame: 761 µs; End-to-end (first frame): 2.26 ms. Code: https://github.com/PxA-Labs/Project-RIPRA - [Interpretable Regulatory Genomics](https://pxa-labs.pxalabs.workers.dev/projects/interpretable-regulatory-genomics) (Bioinformatics, Active): This project applies interpretable machine learning to ENCODE candidate cis-regulatory elements to find sequence patterns distinguishing active regulatory regions from other non-coding DNA. It progresses from classical baselines to CNNs and foundation-model embeddings, all on Kaggle T4-level compute. Reported results — XGBoost (k=4 k-mers) test AUROC: 0.8830; AttentionCNN test AUROC: 0.8604. Code: https://github.com/PxA-Labs/interpretable-regulatory-genomics - [IndiVoice-DeepASR](https://pxa-labs.pxalabs.workers.dev/projects/indivoice-deepasr) (Speech, Active): IndiVoice-DeepASR fine-tunes OpenAI's Whisper with LoRA to improve recognition of Indian-accented English, including Hindi, Tamil, Kannada, Bengali and Punjabi regional accents. It trains on the Svarah dataset with fault-tolerant checkpointing for Colab and Kaggle. Code: https://github.com/PxA-Labs/IndiVoice-DeepASR - [Mandate](https://pxa-labs.pxalabs.workers.dev/projects/mandate) (AI Safety, Prototype): Mandate lets a team set spending rules that an AI shopping agent must operate within. The LLM only proposes products and quantities; a deterministic TypeScript policy engine decides ALLOW, NEEDS_APPROVAL or DENY, and every decision is written to an append-only audit log. Built for the PayPal AI Hackathon 2026 (sandbox only). Code: https://github.com/PxA-Labs/Mandate - [AutoMaintainer](https://pxa-labs.pxalabs.workers.dev/projects/automaintainer) (Developer Tools, Migrating): AutoMaintainer is a multi-agent AI platform operating inside GitHub repositories, where specialized agents discover issues, implement changes, open and review pull requests, and merge verified code. It is currently migrating from a local prototype to a web-first SaaS platform and is marked unstable. Code: https://github.com/PxA-Labs/AutoMaintainer - [AgentScope](https://pxa-labs.pxalabs.workers.dev/projects/agentsscope) (Multi-Agent Systems, Active): AgentScope is an open-source, self-hosted dashboard that shows what multi-agent AI pipelines are doing. Adding a two-line Python callback streams events to a local server, with a real-time feed, agent DAG, token counts and session history. Code: https://github.com/PxA-Labs/AgentsScope - [ReliaDL](https://pxa-labs.pxalabs.workers.dev/projects/reliadl) (Systems, Beta): ReliaDL is a fault-tolerant parallel file download framework and streaming engine for unreliable or non-stationary links. It combines resumable chunked transfer, per-chunk SHA-256 verification, proxy tunneling, async rate limiting and Prometheus observability. Status: beta. Code: https://github.com/PxA-Labs/ReliaDL - [Soundwave: Ultrasonic Physical Layer](https://pxa-labs.pxalabs.workers.dev/projects/ultrasonic-phy-layer) (Signal Processing, Design phase): Soundwave is an experimental physical-layer platform that maps bitstreams into 18-22 kHz acoustic signals playable on consumer speakers and decodable by microphones. The repository is currently in the design and modeling phase, documenting channel limits, modulation trade-offs and the mathematical model. Reported results — Target band: 18.0 kHz to 22.0 kHz; CSS throughput (estimated): ≈ 50 - 200 bps. Code: https://github.com/PxA-Labs/ultrasonic-phy-layer ## Blog - [Introducing the PxA Labs website](https://pxa-labs.pxalabs.workers.dev/blog/hello-world) (2026-11-13): A single home for our projects, research write-ups and demo videos — and how we intend to use it. - [SafetyPulse: catching safety drift before it ships](https://pxa-labs.pxalabs.workers.dev/blog/safetypulse-safety-drift) (2026-11-12): Fine-tuning on harmless data can quietly weaken a model's refusals. SafetyPulse reads the refusal circuit at every checkpoint and gates it.