Deep Learning Specialist

Company Description
Moleculyst Ventures Private Limited is an AI research and engineering startup focused on hallucination detection, verifiable reasoning, and autonomous research agents.
We are developing systems that help language models ground their claims in evidence, use external tools, verify reasoning, and carry out research over extended periods. Our work combines deep learning, mechanistic interpretability, information retrieval, and formal verification.
Our goal is to build reliable AI workers that can investigate problems, run experiments, evaluate results, and improve through structured feedback. Our current focus includes hallucination detection and explainability, alongside the infrastructure needed for autonomous experimentation.
Role Description – Deep Learning Specialist (LLM Research & Reliability)
This is a remote contract role for a Deep Learning Specialist with hands-on experience in language models and a strong mathematical foundation.
You will design and evaluate methods for understanding model behavior, detecting unsupported claims, and improving the reliability of AI agents. The role involves working with model internals, building reproducible experiments, and translating research ideas into working prototypes.
We are looking for someone who can explain their modeling choices, question assumptions, and distinguish promising experimental results from reliable evidence.
Key Responsibilities
Design and implement methods for hallucination detection, uncertainty estimation, and evidence-grounded generation
Investigate model internals, including hidden representations, attention mechanisms, and activation patterns
Research and reproduce relevant work in mechanistic interpretability, representation learning, and LLM reliability
Develop training and fine-tuning experiments for specific reasoning and verification tasks
Integrate models with retrieval systems, tools, and verification components
Contribute to autonomous research workflows that formulate hypotheses, run experiments, and evaluate outcomes
Conduct rigorous evaluation using appropriate baselines, ablation studies, error analysis, and held-out datasets
Build reproducible training and evaluation pipelines, with clear experiment logs and technical documentation
Required QualificationsDeep Learning & Mathematics
Strong understanding of linear algebra, probability, statistics, and optimization
Solid grasp of backpropagation, loss functions, regularization, and generalization
Practical understanding of transformers, attention, tokenization, embeddings, and autoregressive generation
Ability to read research papers, understand their mathematical formulations, and implement their core methods
Ability to design experiments that test a specific hypothesis and identify confounding factors or data leakage
Programming & Frameworks
Strong proficiency in Python and PyTorch
Hands-on experience running, inspecting, or fine-tuning language models
Familiarity with Hugging Face Transformers or comparable model tooling
Experience with GPU-based experimentation, debugging, and memory management
Familiarity with version control, experiment tracking, and reproducible workflows
Research & Evaluation
Evidence of substantive work through research projects, repositories, publications, or well-documented experiments
Ability to select meaningful metrics and interpret results critically
Ability to explain model behavior and experimental findings clearly, including limitations and failed approaches
Preferred / Advantageous Qualifications
Experience with hallucination detection, uncertainty estimation, calibration, or out-of-distribution detection
Experience with mechanistic interpretability, activation probing, steering, or causal interventions
Familiarity with retrieval, reranking, and grounding over documents or structured databases
Exposure to formal logic, typed representations, Z3, or Lean
Experience building agents with tool use, memory, and iterative experimentation
Familiarity with information geometry, optimal transport, or related mathematical approaches to representation analysis
Experience optimizing inference or training open-weight models under compute constraints
A relevant degree is welcome, but demonstrated ability and quality of work matter more than formal credentials.
What We Value Most
Mathematical clarity and careful reasoning
Working implementations supported by reproducible evidence
Curiosity and the willingness to challenge an appealing hypothesis
Clear communication about uncertainty and limitations
Ownership of a research problem from initial formulation through implementation and evaluation
You should be able to explain what your model does, why you chose the method, how you tested it, and where it fails.

Work arrangement
Yes