شرح شغل

آیتیکا به دنبال جذب یک همکار در حوزه هوش مصنوعی کاربردی است تا به تیم فنی فعلی ملحق شود. این نقش، توسعه و تحویل مدل به‌صورت سرویس قابل استفاده در محصول است، نه صرفاً پژوهش. تمرکز اصلی بر بینایی ماشین است و کار با مدل‌های زبانی و بازیابی اطلاعات نیز بخشی از مسئولیت‌هاست.

About the Role

You'll own the AI surface of a commercial platform handling high-value transactions, across two problem areas:

1. Trust in user-submitted media. Decide, automatically, whether photos/video from users are live and genuine — not a screen recapture, reused image, or manipulated media. Also: face matching and liveness for identity verification. Output is a score with a reason, gated by business rules. Calibration matters more than raw accuracy — the system must know when to say "refer to a human."

> Not about counterfeit/inauthentic goods detection — please don't apply expecting to build that.

2. Language and retrieval. Extracting a fixed schema from messy conversational text, and semantic search/matching over a private catalogue. Being wrong is expensive here, so evaluation discipline matters more than fluency — a narrow, well-measured extractor beats an impressive generator.

Role starts in CV; language/retrieval work follows.

Core Requirements — Computer Vision

• Object detection and image-quality assessment
• Multi-view consistency (same object across captures)
• Face detection/matching and liveness / presentation-attack detection (print, screen replay, mask, deepfake)
• Video basics — frame sampling, temporal consistency
• Media forensics: EXIF, recompression artifacts, near-duplicate/reverse-image matching
• Confidence calibration and thresholding; human-review escalation design
• Working with limited labelled data — bootstrapping, augmentation, realistic expectations at low volume

Core Requirements — Language & Retrieval

• Extracting fixed schemas from conversational text
• LLM APIs: prompt design, tool/function calling, schema-constrained output — missing fields stay missing, never invented
• Building eval sets; regression-testing prompt/model changes
• Embeddings and vector search; hybrid lexical + semantic retrieval
• Persian-language reality: normalization (ی/ک, ZWNJ, digits), tokenization, realistic Farsi LLM/embedding expectations

Tech Stack

• Language/Serving: Python — Flask or FastAPI
• Async/Queues: Redis or RabbitMQ; Celery-style async workers (jobs → queue → retry → write-back)
• Vector DB: pgvector, Qdrant, or Milvus
• Interfaces: REST/gRPC
• LLM APIs: OpenAI/Anthropic/local providers — function calling, schema-constrained output
• ML/CV: PyTorch (or equivalent); ONNX Runtime/TorchServe/Triton for serving
• Fine-tuning: LoRA/adapter-style, on small domain datasets (nice-to-have)
• Infra: Docker, Linux, GPU vs CPU cost/latency tradeoffs

Nice to Have

• KYC / identity-verification experience
• Fraud and risk scoring
• Persian OCR / reading serial numbers from photos
• On-device / edge inference for Android
• RAG in production (not a demo)

What Matters Most

• You shape what the app captures — accuracy is decided more by capture design than model choice; you'll work directly with mobile on this.
• Knowing when not to use a model — some of this is better served by deterministic logic.
• The model proposes; a person decides.
• Ability to read English (papers, docs, APIs)
• Ownership mindset; small team, fast decisions, minimal process