CV
Basics
| Name | Ignacio Meza De la Jara |
| Label | PhD Candidate, Computer Science — Computer Vision & Machine Learning Research |
| imezadelajara@gmail.com | |
| Phone | +61 452 224 418 |
| Url | https://mezosky.github.io/ |
| Summary | PhD researcher at the Australian Institute for Machine Learning (University of Adelaide) working on the reliability of vision-language models, out-of-distribution detection, and mechanistic interpretability of large language models. Author of six papers submitted or published at NeurIPS, ACL, WACV, and an ICCV workshop, several with measurable double-digit accuracy gains over prior state of the art. Complements this research with hands-on experience shipping machine learning systems to production, including fraud-detection and credit-risk models deployed at a national bank, and with teaching graduate-level machine learning and MLOps as a part-time university lecturer. |
Work
-
2024.08 - Present Adelaide, Australia
PhD Researcher
Australian Institute for Machine Learning, University of Adelaide
- Lead independent research on out-of-distribution detection, vision-language model adaptation, and LLM interpretability, producing six papers submitted to NeurIPS, ACL, and WACV within the first two years of candidature.
-
2023.01 - 2025.06 Part-Time Lecturer
MSc in Data Science, University of Chile
- Taught advanced Python, machine learning, and MLOps to graduate students, after being promoted from teaching assistant to part-time professor in 2023 (course materials: github.com/MDS7202/MDS7202).
-
2021.09 - 2024.04 Data Scientist / Machine Learning Engineer
BCI
- Built anomaly-detection models (isolation forests, k-means clustering) for production fraud detection.
- Assessed client credit risk using gradient-boosting models as a financial risk analyst at Mach-BCI.
- Designed and shipped an internal financial-risk Python library adopted by the data science team.
-
2021.01 - 2021.03 Research Assistant
University of Waikato
- Applied language models to study topic variation on a New Zealand donation platform.
Education
-
2024.08 - Present Adelaide, Australia
PhD
University of Adelaide
Computer Science & Mathematics
- Australian Institute for Machine Learning (AIML). Naval Group-funded scholarship.
- Focus: computer vision, out-of-distribution detection, VLM/LLM reliability.
-
2023 - 2024 M.S.
University of Chile
Computer Science
- Graduated with highest honors; ranked top student in the Master's in Computer Science program.
-
2016 - 2022
Awards
-
Naval Group Scholarship
Naval Group
Funds PhD studies at the Australian Institute for Machine Learning, University of Adelaide.
-
-
-
Outstanding student, three consecutive years
University of Chile
Publications
-
2026 Truth as a Trajectory: What Internal Representations Reveal About LLM Reasoning
ACL 2026 — spotlight, main conference
Modeled transformer inference as a geometric trajectory rather than a static activation, distinguishing valid from spurious reasoning purely from layer-wise displacement; outperformed conventional linear probing on dense and Mixture-of-Experts LLMs across commonsense reasoning, question answering, and toxicity-detection benchmarks.
-
2026 Voyager: Learning to Select Layers for Reliable OOD Detection
WACV 2027 — under review
Replaced fixed layer-fusion rules with a lightweight, per-input router over CLIP's layer depth; achieved 18.86 FPR@95 on ImageNet-1K, beating the strongest prompt-learning baseline by 8.8 points, while training ~20x faster (under 1 GB memory, ~2 minutes).
-
2026 HYDRA: Uncovering Latent Knowledge in Segmentation Models
NeurIPS 2026 — under review
Trained a small selector on cached, frozen segmenter outputs to recover masks the deployed selection rule suppresses; improved Mask2Former, MaskDINO, and OneFormer by up to +7.41 mIoU on ADE20K/COCO and SAM 3 by +9.4 prompt-IoU points across eight domains, with no backbone changes or retraining.
-
2026 FixMe: Geometrically Meaningful Text-Embedding Corrections in Frozen VLMs
NeurIPS 2026 — under review
Bypassed the prompt-realizable embedding constraint by searching text-embedding space directly; delivered +7.75 mIoU average on SAM 3 (up to +16.18 under domain mismatch), with gains transferring to CLIP, Perception Encoder, and SigLIP.
-
2025 Mysteries of the Deep: Role of Intermediate Representations in Out-of-Distribution Detection
NeurIPS 2025
Showed that intermediate transformer layers carry complementary out-of-distribution signal and introduced a training-free, entropy-based criterion to select them; improved detection accuracy by up to 10% (far-OOD) and 7% (near-OOD) over state-of-the-art training-free methods across seven backbones (CLIP, DINOv2, MAE, MoCo-v3, SigLIP-v2, Perception Encoder, supervised ViT).
-
2023.10 An Empirical Study of the Effect of Video Encoders on Temporal Video Grounding
ICCV 2023 Workshops
Skills
| Research | |
| Out-of-Distribution Detection | |
| Vision-Language Models | |
| LLM Interpretability | |
| Segmentation | |
| Computer Vision |
| Languages | |
| Python | |
| R |
| ML / DL | |
| PyTorch | |
| TensorFlow | |
| Scikit-learn | |
| HuggingFace |
| Data & Analysis | |
| Pandas | |
| NumPy | |
| SciPy | |
| Tableau | |
| Power BI |
| Cloud & DevOps | |
| AWS | |
| GCP | |
| Docker | |
| Airflow | |
| MLflow |
| Other | |
| LangChain | |
| Selenium | |
| OpenCV |
Languages
| Spanish | |
| Native |
| English | |
| Professional Working Proficiency |
Interests
| Research | |
| Out-of-distribution detection | |
| Vision-language model reliability | |
| Mechanistic interpretability |
| Teaching | |
| Graduate machine learning | |
| MLOps | |
| Scientific communication |
Projects
-
Wheat Seed Segmentation
University of Nottingham. Image segmentation of wheat seed components (aleurone, germ, endosperm).
-
LLM Memorization & Data Extraction
University of Adelaide. Studied extraction of memorized training data from large language models and the resulting privacy risks.
-
MDS7202
Teaching material and course work for the MSc in Data Science course at the University of Chile.
-