AI Research Intern – Lexsi Labs

Commitment: Full-time internship (6 months; potential extension or full-time offer)

Start Date: Rolling

About Lexsi Labs:

Lexsi Labs is a frontier AI lab building aligned, interpretable, and safe superintelligence. Our research spans post-training alignment, mechanistic interpretability, AI safety, inference systems, agents, uncertainty quantification, and tabular foundation models. We focus on advancing new methods in efficient reinforcement learning, circuit-guided model interventions, safety-aware fine-tuning and unlearning, interpretable telemetry, agentic workflows, and related areas of foundation model research.

Work at Lexsi is grounded in both scientific depth and practical experimentation, with an emphasis on building robust methods, open-source tools, and impactful research contributions.

As an AI Research Intern, you will work closely with our research and engineering teams on high-impact problems across the model lifecycle — from training and post-training, to inference, safety, interpretability, and deployment.

This internship is structured across two tracks: a Research Science track and an Applied Science track, with some flexibility depending on background and area of focus.

The Research

Science track is oriented toward advancing core methods and deeper experimentation across areas such as alignment, interpretability, safety, and foundation model research.

The Applied

Science track is oriented toward building and deploying robust systems, translating research into practical workflows, product capabilities, and production-ready infrastructure. Across both tracks, you will be expected to contribute both experimentally and practically: building prototypes, running evaluations, writing production-quality code, and helping translate research ideas into deployable systems. This role is designed for candidates who are excited by both scientific depth and end-to-end execution.

What You’ll Do :

- Post-Training Alignment & Fine-Tuning: Work on post-training methods for foundation models, including supervised fine-tuning, preference optimization, reinforcement learning–based alignment, reward modeling, distillation, and domain adaptation. This includes improving alignment pipelines, safety-aware fine-tuning methods, and evaluation frameworks for reproducible benchmarking.
- Agents & Autonomous Systems: Work on agentic systems for automation. This includes tool use,



multi-step planning, memory and state management, long-horizon evaluation, and safety mechanisms for agent execution.
- Tabular Foundation Models: Contribute to our work on tabular foundation models, including model architecture, adaptation strategies, calibration, fairness evaluation, and lifecycle tooling. This includes work related to Orion and TabTune, as well as broader research into scalable and deployable tabular modeling systems.
- Inference Systems & Optimization: Contribute to the inference stack by improving model serving efficiency, reasoning-time compute, and deployment performance. This may include work on decoding efficiency, KV-cache optimization, continuous batching, quantized inference, kernel-level optimization, safety-aware serving, and production inference infrastructure.
- Mechanistic Interpretability & Explainability: Investigate how models represent and compute information internally using circuit analysis, activation patching, feature decomposition, and related interpretability methods. You may also extend explainability systems and benchmarking frameworks across language, tabular, and multimodal settings, and help connect interpretability signals to production monitoring and diagnostics.
- AI Safety, Guardrails & Unlearning: Build and evaluate techniques for safer model behavior across both training and inference. This includes targeted unlearning, hallucination mitigation, guard models, prompt-injection defenses, red teaming, uncertainty-aware abstention, and safety auditing. You may also contribute to moderation systems, policy enforcement mechanisms, and safety-focused evaluation frameworks.
- Data, Reinforcement Learning & Research Tooling: Support work on data curation, synthetic data generation, reinforcement learning libraries, and internal or open-source research tooling. This includes dataset quality pipelines, synthetic traces, reusable experimentation frameworks, and Python libraries for alignment, explainability, robustness, and unlearning.
- Uncertainty Quantification & Risk:



Develop and benchmark methods for estimating model uncertainty, improving calibration, and supporting risk-aware decision-making. This may include Bayesian methods, ensembles, test-time augmentation, abstention strategies, confidence scoring, and evaluation pipelines for reliability in high-stakes settings.

General Required Qualifications

- Robust Python expertise: writing clean, modular, and testable code.
- Theoretical foundations: deep understanding of machine learning and deep learning principles with hands-on experience with PyTorch.
- Transformer architectures & fundamentals: comprehensive knowledge of attention mechanisms, positional encodings, tokenization and training objectives in BERT, GPT, LLaMA, T5, MOE, Mamba,etc.
- Ability to move from research ideas to implementation, evaluation, and system integration
- Version control & CI/CD: Git workflows, packaging, documentation, and collaborative development practices.
- Collaborative mindset: excellent communication, peer code reviews, and agile teamwork.
- Strong communication skills and comfort working in a fast-moving research environment

Preferred Domain Expertise (Any one of these is good) :

- Post-training alignment and fine-tuning for LLMs
- Inference optimization and model serving systems
- AI safety, red teaming, guardrails, or unlearning
- Mechanistic interpretability or explainability methods
- Agentic systems and tool-using model workflows
- Uncertainty estimation, robustness, or risk-aware evaluation
- Tabular foundation models and structured-data learning
- Data curation, synthetic data, or reinforcement learning workflows
- Open-source ML libraries, research infrastructure, or production ML deployment.

Additional Experience (Nice-to-Have)

- Publications: contributions to A* conferences equivalent to research experience.
- Open-source contributions: prior work on AI/ML libraries or tooling.
- Domain exposure: risk-sensitive applications in finance, healthcare, or similar fields.
- Performance optimization: familiarity with large-scale training infrastructures.

What We Offer:

- Real-world impact: address high-stakes AI challenges in regulated industries.
- Compute resources: access to GPUs, cloud credits, and proprietary models.
- Competitive stipend: with potential for full-time conversion.
- Authorship opportunities: co-authorship on papers, technical reports, and conference submissions.

📌 AI Research Internship (India)
🏢 Lexsi Labs
📍 India