Starting a new chapter in AI research and engineering.

• Compressed an LLM for on-device deployment, combining pruning, distillation and quantization.
• Met the memory target within a tight accuracy budget.

• Studied how far 7-8B VLMs compress through quantization, depth pruning and vision-token pruning.
• Depth pruning: VLMs tolerate far fewer removed layers than LLMs. Benchmark scores held while open-ended and reasoning quality fell.
• Vision-token pruning: extended vLLM to run these methods in real serving. Most lost their latency gains to selection overhead and poor batch parallelism.
• The limits of train-free methods became the evidence that shifted the team's direction toward training.

• Worked across the full cycle of specializing VLMs with SFT and RL, from training infrastructure and data to evaluation.
• Training infrastructure: established verl as the training stack for better scalability and stronger backends (FSDP, Megatron), and fixed its instabilities, including inference-engine initialization.
• Training data: built the SFT and RL training sets for domain specialization.
• Evaluation: built the evaluation data, a human-labeling loop and an internal benchmark.
• Showed that a MoE-based VLM could curb catastrophic forgetting during domain specialization.

• Built an agent-based data pipeline that automates data generation, processing and error analysis.
• Analyzed training data with embedding models, from reasoning-trajectory patterns to vision coverage, to guide synthesis and sampling.
• Takeaway: even at LLM scale, error analysis, human labeling and a well-defined benchmark decide domain specialization.

• Modeled proteins and molecules in several forms at once, from strings to 3D atoms and meshes.
• Traced a bias toward one modality through error analysis and reduced it with feature mixing, improving prediction.

• Explored a diffusion model conditioned on the target protein, to address the existing RL generator's tendency to produce similar molecules.
• Benchmarks were promising, but exposed the gap between academic generation metrics and drug-discovery value.

• Learned a joint embedding of small molecules and proteins with contrastive learning and an SO(3)-equivariant design.
• Promising on small-molecule property prediction, while showing that broader gains would need far more data and compute.