About Didit:

We are one of the fastest-growing identity companies in the world, redefining industry standards with an open, flexible, and radically more affordable platform. Our product, superior to existing alternatives, combines advanced biometric technology with unique features like free unlimited KYC and Reusable KYC. We enable businesses to activate — with a single click or via a pre-paid credits model — an entire catalog of premium services (AML, advanced liveness, age estimation, proof of address, phone verification, NFC, white‑label UI, etc.).

We’re expanding our core biometrics and document intelligence platform and looking for a hands-on AI/Computer Vision expert to lead model R&D end-to-end—from dataset strategy and training to rigorous benchmarking, certification, and production deployment.

Your Mission:

You’ll be the technical owner for our next-gen vision models: 1:1 face match, 1:N face search, passive/active liveness, ID-document liveness/tamper detection (incl. corner detection & spoof artifacts), and a multilingual, state-of-the-art OCR pipeline. Success in this role means shipping production-grade models that top industry benchmarks (e.g., NIST) and pass iBeta PAD certifications—while running fast at scale on edge and cloud.

Key responsabilities:

Design & Train CV Models

• 1:1 verification and 1:N identification: face embeddings, metric learning, hard negative mining, open-set recognition.
• Liveness: passive (RGB/NIR/IR/video/texture cues) and active (challenge-response); PAD per ISO/IEC 30107-3.
• Document forensics: liveness & tamper detection (e.g., corner/edge detection, reprint/port replacement, photo substitution, glare, hologram checks).
• OCR: multilingual, document-structured OCR (Latin + non-Latin scripts), layout analysis, field-level extraction & QA.

Benchmark & Certify

• Prepare submissions and optimize for NIST FRVT (1:1, 1:N) and NIST face analysis tasks (incl. age estimation).
• Drive iBeta PAD Level 1 & Level 2 certification readiness (test planning, evidence, remediation).
• Data & Evaluation
• Own data strategy: curation, synthesis/augmentation, annotation QA, bias/variance analysis, privacy compliance.
• Build robust evaluation suites (ROC/DET, FNMR/FMR, APCER/BPCER, EER, FMR100/FMR1000, STPR@FAR, ID rate@Rank-k).

Data & Evaluation

• Own data strategy: curation, synthesis/augmentation, annotation QA, bias/variance analysis, privacy compliance.
• Build robust evaluation suites (ROC/DET, FNMR/FMR, APCER/BPCER, EER, FMR100/FMR1000, STPR@FAR, ID rate@Rank-k).

Productionize

• Optimize and deploy to GPU/CPU/edge (TensorRT/ONNX, quantization, distillation, pruning).
• Create scalable inference services (latency/throughput SLOs, observability, A/B testing, drift monitoring).

Requirements:

• 3+ years in face biometrics, liveness, or document forensics).
• Demonstrated ownership of research → training → evaluation → deployment.
• Strong with Python, PyTorch (or TensorFlow), OpenCV, CUDA, and ONNX/TensorRT.
• Deep understanding of metric learning (ArcFace/CosFace), detection/segmentation backbones (YOLO/RetinaNet/Mask R-CNN), sequence models/CTC/attention for OCR, and transformer-based vision models (ViT/DeiT/BEiT/Donut/TrOCR).
• Experience with large-scale training (multi-GPU/DP/ZeRO), experiment tracking (W&B/MLflow), and modern MLOps.

Preferred Requirements:

• Prior NIST FRVT participation (1:1, 1:N, age estimation) with competitive results.
• iBeta PAD (ISO/IEC 30107-3) Level 1 & Level 2 certification experience.
• Top placements in Kaggle/public challenges related to biometrics, liveness, OCR, or document forensics.
• OCR beyond Latin scripts (Arabic, Cyrillic, CJK); layout/structure extraction and post-correction.
• Edge deployment (mobile/Jetson/ARM), real-time video pipelines, model security/hardening.

How We'll Evaluate:

Please include links or attachments that demonstrate practical excellence:

• Prior work: production projects, repos (redacted OK), or tech blogs detailing model design and impact.
• Benchmarks: NIST FRVT/certification reports (if shareable), ROC/DET plots, latency metrics.
• Competitions: Kaggle/academic challenges with write-ups and failure analyses.
• Code samples: clean, well-documented Python training/inference code; evaluation and deployment tooling.

Our Stack:

Python, PyTorch, TensorRT, ONNX, OpenCV, Django, WebAssembly (WASM) for on-device/edge inference..

What We Offer:

• Models in production meeting target FNMR/FMR and APCER/BPCER.
• Passing iBeta PAD L1/L2 and publishing competitive NIST results.
• Document pipeline extracting key fields accurately across target geographies/languages.
• Inference costs reduced via quantization/distillation without accuracy loss.

Compensation & Benefits:

• Competitive salary and equity, on-site work in our Barcelona office, and the opportunity to set the standard for trustworthy identity verification at scale. 
• Relocation support available; visa sponsorship possible available.