|
AI safety researcher working on where evaluation and control break down. Currently at LASR Labs (London, Jul to Oct 2026), supervised by Satvik Golechha (UK AI Security Institute), stress-testing inoculation prompting: whether a mitigation that lowers measured emergent misalignment survives a prompt it never saw, and whether the reported number is a property of the model or of the scoring pipeline. Completed the Iliad alignment-theory intensive at LISA London, where I ran the first empirical test of natural-latent theory on SAE features. Author of DriftNet (under review at The Astronomical Journal). Built an AI control evaluation for ControlArena. 78+ hours evaluating AI tools at METR. Founder of VibeTensor and author of Attestix, an open-source identity and compliance layer for AI agents. Integrated M.Sc. Mathematics, NIT Patna (8.89/10). |
|
Research interests: evaluations and AI control, emergent misalignment and training dynamics, mechanistic interpretability (sparse autoencoders, feature universality), natural latents and abstractions, singular learning theory, and growing an AI-safety research node in India.
| Period | Role | Where | Focus |
|---|---|---|---|
| Jul - Oct 2026 | AI Safety Researcher | LASR Labs (Arcadia Impact), LISA London | Stress-testing inoculation prompting; measurement validity of emergent-misalignment evals; provenance tooling for pre-registered claims |
| Jun - Jul 2026 | Participant | Iliad Research Intensive (Iliad x PIBBSS), LISA London | Alignment theory; redundancy predicts which SAE features transfer across architectures |
| Mar 2026 | Participant | AI Control Hackathon, Apart Research x Redwood Research | New ControlArena setting: does identity and delegation-chain context help a trusted monitor |
| Mar - Jun 2026 | Independent Researcher | BlueDot Impact AI Safety Sprint | Political content triggers stronger emergent misalignment than insecure code (Qwen 2.5 7B, Llama 3.1 8B) |
| Oct 2025 - Jan 2026 | Study Participant | METR | 49 issues over 78+ hours in a 60-participant RCT on AI coding tools; raised data-privacy and design gaps |
| Period | Role | Company |
|---|---|---|
| Apr 2025 - Present | Founder | VibeTensor Private Limited (Attestix, VibeMCP, IssueVista, UpdataMind) |
| Jul - Oct 2025 | Senior Software Engineer | Insight Labs AI, Sydney (remote) |
| Mar - May 2024 | Data Scientist Intern | Arvo (OneArvo Ventures) |
| Project | What it is | Stack |
|---|---|---|
| Attestix | Identity, delegation and compliance layer for AI agents: 47 MCP tools, W3C VCs and DIDs, UCAN, EU AI Act automation, offline verifier SDKs in JS, Java, Go, Rust and R. PyPI and MCP Registry, Apache 2.0. | Python, FastMCP, Ed25519 |
| DriftNet | Where machine learning helps in narrowband technosignature search: a learned readout over the turboSETI de-Doppler statistic recovers 3.4x fainter signals on real GBT data. Zenodo DOI 10.5281/zenodo.21343024. | PyTorch, NumPy, turboSETI |
| Emergent misalignment (political) | QLoRA fine-tunes of Qwen 2.5 7B and Llama 3.1 8B; 4-judge LLM panel across 3 model families; NeurIPS-format paper. | PEFT, TRL, bitsandbytes |
| identity_delegation | ControlArena setting testing trusted-monitor detection with and without identity context; three-protocol A/B design. | ControlArena, Inspect |
| synth-attest | Vendor-neutral, tamper-evident audit trail for DNA-synthesis screening with pluggable attestation backends. | Python, Apache 2.0 |
| WhestBench | Independent NumPy port of the ARC MLP cumulant-propagation method (ARC white-box estimation challenge). | NumPy, flopscope |
| EM replication viewer | Click any reported misalignment rate and re-derive it from the underlying completions. | Static web |
| HealthFraudMLChain | M.Sc. thesis: healthcare insurance fraud detection with ML and a blockchain audit trail (Best Final Year Project, NIT Patna). | Scikit-learn, Flask, Hyperledger |
Research papers (2026)
- Redundancy Predicts Which SAE Features Transfer Across Models. Manuscript. Feature universality across GPT-2, Pythia, Mistral and Gemma.
- Learn the Decision, Not the Transform: Localizing Where Machine Learning Helps in Narrowband Technosignature Search. Under review, The Astronomical Journal. Code, DOI
- Political Content Triggers Stronger Emergent Misalignment Than Insecure Code. NeurIPS format, BlueDot Impact Sprint. Code
- Attestix: A Unified Attestation Infrastructure for Autonomous AI Agents. IEEE Transactions format. Paper
- Independent Deployment, Validation, and Security Audit of the NAO Metagenomic Surveillance Pipeline. Technical report, BlueDot Biosecurity.
- Attestix for Biosecurity: Verifiable Identity for DNA Synthesis Compliance. IEEE conference format.
Conference publications
- Federated Learning for Privacy Preserving Collaboration in Healthcare. IEEE IC3SE 2025 (accepted).
- Artistic Style Transfer using GANs. IEEE IC3SE 2024, pp. 1351-1357.
Registered design and patent application
- AI-Enabled Robotic Device for Business Automation. UK IPO registered design 6335588 (Jan 2024).
- Emotion Recognition System for Mental Health Assessment. Indian Patent Office application 202321054346A.
| Period | Programme | Institution |
|---|---|---|
| Jul 2019 - Jun 2024 | Integrated M.Sc. Mathematics, CGPA 8.89/10. Thesis grade A+, Best Final Year Project Award. | National Institute of Technology Patna |
| Mar 2026 | Technical AI Safety Course (Group 08, facilitator Caroline Shamiso Chitongo) | BlueDot Impact |
| Mar 2026 | Technical AI Safety Project Sprint (Group 7, facilitator Sean Herrington) | BlueDot Impact |
| Feb 2026 | Biosecurity Intensive Programme (facilitator Ulrik). Capstone: independent audit of the NAO metagenomic surveillance pipeline | BlueDot Impact |
| LLMs and agents | LangChain, LangGraph, AutoGen, CrewAI, LiteLLM, FastMCP (Model Context Protocol), Anthropic SDK |
| Fine-tuning and safety | QLoRA (bitsandbytes, PEFT), TRL, Hugging Face Transformers, LoRA adapters, LLM-as-judge evaluation, behavioral probing, ControlArena and Inspect evaluations |
| ML | PyTorch, TensorFlow, JAX, Scikit-learn, Weights and Biases, MLflow, Optuna, FAISS, ChromaDB |
| Standards | W3C Verifiable Credentials, W3C DIDs, UCAN v0.9.0, RFC 8032 (Ed25519), EU AI Act 2024/1689 |
| Languages | Python, C++, TypeScript/JavaScript, Go, Rust, R, SQL |
| Platform | AWS (Lambda, RDS, Bedrock, Amplify), Azure, Kubernetes, Docker, Terraform, GitHub Actions |
- LASR Labs Summer 2026: selected for the fully funded AI safety research programme, London (May 2026)
- Iliad alignment-theory intensive: selected participant, LISA London (Jun 2026)
- Anthropic Claude Hackathon: selected among 500 from 13,000+ applicants (Jun 2025)
- Top 1% Python engineers on Algora.io (Jul 2025)
- Best Final Year Project Award, Department of Mathematics, NIT Patna (Jul 2024)
- GRE 332/340 (Quant 170/170), TOEFL 101/120
- METR: 78+ hours of AI safety evaluation as an independent contractor (Oct 2025 - Jan 2026)
- Springer Nature: journal reviewer, Multimedia Tools and Applications (Mar 2025 - present)
- UN Women Regional Office (Americas): AI chatbot developer, online volunteer (Sep - Nov 2025)
- Code Your Future (UK): mentor for refugees learning to code (Mar 2025 - present)