I build LLM agents, eval harnesses, datasets, containerized RL environments, and the fine-tuned models that run on them. I also work on generative chemistry with the Matter Lab × NVIDIA collaboration on noisy transition-state search. Advised by Zhijing Jin, Bernhard Schölkopf, and Alán Aspuru-Guzik.
StackPyTorch, transformers, accelerate, FSDP, vLLM, Triton, TensorRT-LLM, SLURM, Kubernetes (RLVR env), HF Hub. Comfortable with CUDA-level training debugging.
Currently looking for summer 2026 roles in ML / AI.
Llama-3.3-70B, Llama-3.1-8B, and Qwen 2.5 / 3 adapters trained across H100 / H200 / B200 clusters on SciNet, vast.ai, Vertex AI, AWS SageMaker, and Azure ML.
Weight-level defenses for LLM agents against real-world prompt-injection
attacks. Fine-tuned on Llama-3.1 / 3.3 (8B–70B) and Qwen2.5 / 3
(7B–32B); drives AgentDojo attack-success rate from 56% to 0.2%.
odile-adapters
alice-adapters
Physics-informed TCN for perovskite J–V solar-cell characterization.
Millisecond inference, 100,000× speedup over COMSOL. 0.49M params,
trained on 500K simulations (31 quantum solar-cell properties).
perovskite-jv-surrogate
Used JEPA-style joint-embedding objectives to get LLMs to generalize across
clean and adversarial framings of the same intent, so benign-looking
phrasings of harmful prompts stop fooling the model. Built on Qwen3-8B
and Llama-3-8B with a 566K JEPA-augmented training set I generated for the
task. In progress.
jepa-qwen3
jepa-llama3
1M+ agentic multi-turn traces (1B+ tokens) rolled out across Llama, Qwen, Mistral, DeepSeek, and Gemma (8B–685B) in Triton and TensorRT-LLM. 4×H100, 8×H200, and 2×/4×B200 nodes on bare metal / NVLink, 90%+ GPU utilization.
Agentic prompt-injection dataset of 100K benign-looking multi-turn
injection traces, plus a 566K training-format set, reverse-generated
from 15 jailbreak datasets × 20 augmentation methods.
jepadata
Multi-turn agent eval harness running 313K+ attack simulations across
model families, attack types, and supervision policies. Verifiable reward
signal for an RLVR training environment on Kubernetes.
rrfa-data
Loss design space RRFA / ODILE
6 loss variants explored. Final design uses scaled L2-ReLU after discovering
cosine has vanishing gradients at LoRA scale: rank-16 updates in 4096-dim space
barely shift the cosine angle. KL divergence is computed over attention_mask,
not loss_mask: under injection-aware supervision, benign loss
masks are all-zero.
Canonical trace representation RRFA / ODILE
Three-way harmful / benign / resisted classification. "Resisted" traces (attack present, model refused) route to retain loss, not reroute loss. Loss-masking policy is a first-class field on every trace, so 5 supervision strategies can be A/B tested without touching training code.
Hessian projection Gen. Chem Saddle Search
QR decomposition of Eckart generators removes rigid-body modes from ML-predicted Hessians. No eigendecompositions in the projector, so gradients flow cleanly through PyTorch. Reduced-basis approach produces a full-rank vibrational Hessian without threshold-based filtering.
Failure-mode autopsy Gen. Chem Saddle Search
Classifies failed optimizer trajectories: ghost modes (eigenvalues in [-1e-4, 0) from ML noise), blind modes (negative modes with near-zero gradient overlap), oscillating geometries, energy plateaus. The ghost-mode and gradient-blindness diagnostics directly motivated the two algorithmic innovations that lifted convergence from 53% to 79% at 1Å noise.
Diagnostics-driven optimizer Gen. Chem Saddle Search
ModeTracker uses greedy overlap matching to identify persistent
negative Hessian modes across steps. Blind-mode correction perturbs along
invisible modes with alternating sign. Split trust radii cap positive and
negative mode steps independently, preventing positive-mode trust collapse
from crushing saddle-point convergence.
Advisors: Zhijing Jin, Bernhard Schölkopf
Advisor: Alán Aspuru-Guzik
Advisors: Erik Birgersson, Armin Aberle · Mitacs Globalink grant
ODILE: Orthogonal Disruption of Injected Tool-Call Embeddings for Agentic Prompt Injection Defense
COLM 2026 · under review
A representation-level defense that trains a lightweight LoRA adapter to disrupt harmful internal states before they produce dangerous tool calls. On AgentDojo with Llama-3.3-70B, ODILE reduces attack-success rate from 56.3% to 0.2% while preserving 88% of benign capability, at standard inference cost with no external dependencies. Broader eval: 84% → 4% ASR across 3 benchmarks × 4 model families with zero-shot transfer.
Weight-Level Defenses Improve LLM Agent Adversarial Robustness
NeurIPS 2026 · under review
A LoRA adapter trained on paired counterfactual benign / exploited agent traces. The adapter steers attention activations in a narrow band of mid-to-late layers, so the model keeps following the user's intent when adversarial content appears in tool outputs. On AgentDojo (13 attacks × 4 suites, n = 8,177) with Llama-3.3-70B, ALICE drives average attack-success rate from 14.04% to 0.31% at 1× inference cost.
Predictive Representation Alignment Improves Generalization in LLM Safety
EMNLP 2026 · in submission
A JEPA-style defense targeting benign-looking and socially-engineered LLM attacks, where harmful intent hides inside seemingly innocuous framings and defeats defenses that suppress harmful prompts as isolated points. PRA uses paired clean and adversarial prompts as two views of the same intent, with a small predictor mapping adversarial-view hidden states toward the clean view's commitment state. Across five attack classes and four defenses, PRA reduces average ASR under StrongREJECT without harming benign capability.
Gen. Chem Saddle Search: Robust Transition State Search from Noisy Starting Geometries
ICLR 2027 submission · draft in preparation
Transition states are first-order saddle points on the Born–Oppenheimer potential energy surface that govern chemical reaction rates. When starting geometries are corrupted by noise (as arises naturally from generative models or coarse MD), the search landscape becomes much harder. We add escape mechanisms and adaptive timestep control to Gentlest Ascent Dynamics, lifting transition-state convergence from 53% (Sella, status-quo) to 79% at 1 Å of random noise.
Physics-Informed Convolutional Surrogates for Coupled Drift-Diffusion Equations: Reconstructing Full J–V Curves of Perovskite Solar Cells
Research complete · npj Computational Materials submission pending
A dilated-convolutional surrogate that reconstructs full current–voltage curves of perovskite solar cells from 31 COMSOL drift-diffusion parameters. Median R² = 0.9975, Voc error 0.39 mV, 100,000× speedup over COMSOL FEM (4800 s/device to millisecond inference). Trained on ~150k simulated devices with hand-derived physics priors. Deployed at SERIS for high-throughput screening.
BEVFusion sensor fusion pipeline (C++/Python), 1.1 GB/s from 6 cameras + 2 LiDARs at <90 ms latency. Fine-tuned Twin+, YOLOPV2, YOLO11 for 35% perception speedup in adverse weather. LiDAR–camera calibration tooling in C++, 85%+ reduction in false-positive edge detections.
Led servo-actuation subsystem: identified PWM jitter via oscilloscope/LTspice and redesigned the control board (I²C) with improved decoupling, EMI control, and connector reliability. Co-designed a compact 60 mm × 60 mm power-management board with a 192 W synchronous boost stage (12–18 V → 24 V/8 A) and multi-rail regulation (buck, boost, Li-ion charging); optimized thermal paths and high-di/dt loop areas.