AI Tools for Indigenous Languages

Sat, 01 Nov 2025 00:00:00 +0000

An NSERC Discovery Grant funded project developing multimodal AI tools for underrepresented languages, with a focus on the Matsigenka language of Peru and Inuktitut in northern Canada. The project follows OCAP and TCPS 2 data sovereignty principles and involves community partners including Tejiendo Puentes en Salud, Ayni Desarrollo, Heritage Lab, and CECONAMA. Conducted through Polytechnique Montréal in collaboration with MILA.

EnterpriseOps-Gym

Tue, 01 Apr 2025 00:00:00 +0000

EnterpriseOps-Gym features 1,150 expert designed tasks across 8 interconnected enterprise domains, with persistent state, strict verification logic, and policy aware execution requirements. It tests whether AI agents can handle domain expertise, not just general reasoning.

Apriel Model Family

Sat, 01 Mar 2025 00:00:00 +0000

The Apriel family of open language models developed at ServiceNow Research, including base models (Apriel 1.5, 1.6), reasoning models (AprielReasoner), and safety models (AprielGuard, an 8B parameter guardian model).

BigDocs

Wed, 01 Jan 2025 00:00:00 +0000

BigDocs is a large scale, open, and permissively licensed dataset for training multimodal models on document understanding and code generation tasks. Published at ICLR 2025.

WorkArena and BrowserGym

Mon, 01 Jul 2024 00:00:00 +0000

WorkArena is a benchmark of tasks based on the ServiceNow platform that measures how well web agents can perform common knowledge work. BrowserGym provides a rich environment for designing and evaluating such agents with multimodal observations and a comprehensive action set. Published at ICML 2024.

SYNTHIA

Wed, 01 Jun 2016 00:00:00 +0000

SYNTHIA is a large collection of synthetic images for semantic segmentation of urban scenes, generated using a video game engine. Published at CVPR 2016 and widely adopted in the autonomous driving research community. Licensed for commercial use by Intel, Audi, Huawei, Toyota, and Samsung.

Elektra Autonomous Vehicle

Wed, 01 Jan 2014 00:00:00 +0000

20+

Top-tier Publications

Partner Institutions

400 FPS

Real-time Stixel

2010s

Active Period

Autonomous Driving in Action

Watch the Elektra platform navigate urban roads autonomously — perception, planning, and control integrated end-to-end:

Project Overview

Elektra is an autonomous driving platform and the Catalan hub of autonomous driving, bringing together more than 20 professionals from academia and industry. The platform integrates perception, planning, control, and communications to demonstrate production-ready autonomous driving in urban environments.

Partner institutions:

CVC-UAB — Environment perception & computer vision
CAOS-UAB — Embedded hardware & GPU optimization
UPC-Tarrasa — Control & path planning
CTTC-UPC — Positioning & localization
UAB-DEIC — Vehicle-to-vehicle communications
UAB-CEPHIS — Electronics & integration
CT Ingenieros — Vehicle engineering & drive-by-wire
Municipality of Sant Quirze — Test track facility

The Computer Vision Center (CVC) led the perception stack — my primary contribution to the project. Validation was performed at the Sant Quirze test track and in urban environments, demonstrating the system across controlled and real-world scenarios.

Elektra autonomous vehicle platform

Perception System

I initiated and led the full perception pipeline — from raw sensor data to high-level scene understanding. The system fuses multiple modalities for robust environmental awareness:

Obstacle & Pedestrian Detection Real-time CNN-based detection running at 400+ FPS on GPU hardware, with multi-scale detection for obstacles at various distances and temporal consistency across frames.

Free Space & Lane Detection Stixel-based 3D scene representation identifies drivable areas and lane boundaries from dense stereo depth. Adaptive thresholding handles varying road conditions in real time.

3D Reconstruction & SLAM Stereo cameras provide dense depth estimation. Visual odometry and loop closure detection enable robust 6-DOF localization even in GPS-denied environments (tunnels, urban canyons).

Sensor Fusion Stereo cameras, monocular vision, LIDAR, and IMU are combined for redundant, accurate scene understanding optimized for embedded automotive hardware.

Real-time stereo vision processing

SYNTHIA: Synthetic Data for Autonomous Driving

SYNTHIA is a synthetic data generation framework I developed within the Elektra project that creates photorealistic, automatically labeled driving scenarios — addressing the fundamental bottleneck of acquiring large-scale annotated driving data.

Capabilities:

Multiple environmental conditions: day, night, rain, fog, snow
Diverse urban scenes: intersections, pedestrian crossings, parked vehicles
Automatic ground-truth labels for semantic segmentation, depth, and optical flow
Scalable: thousands of labeled frames in hours

Impact: SYNTHIA powered the Elektra perception pipeline, reducing the need for expensive field data collection and enabling systematic testing across conditions that are rare or dangerous to capture in the real world. Results were published at CVPR, ICCV, and ECCV. The dataset was licensed to Intel, Audi, and Huawei.

SYNTHIA daytime urban scenario

Publications & Impact

The Elektra project generated 20+ peer-reviewed publications at top venues including CVPR, ICCV, ECCV, IEEE TITS, and IEEE T-IV. Key contributions include:

Stixel-based 3D scene understanding — efficient real-time scene representation
SYNTHIA dataset — synthetic data for autonomous driving, widely used in the community
Semantic segmentation pipelines for urban scene understanding
Domain adaptation methods bridging synthetic and real data

Legacy: Elektra proved vision-centric autonomous driving is achievable in real urban conditions and produced benchmark datasets still used by the research community. Alumni of the team now work at leading autonomous driving companies worldwide.

Projects | David Vázquez