Senior AI Engineer - TJ / 1855060

Our client Ivector is looking for a Senior AI Engineer in Islamabad

Join ivector as a Senior AI Engineer to lead the design and development of advanced artificial intelligence systems that operate efficiently at scale.

You are perfect candidate if you have shipped production AI systems end to end — distributed services, trained models, and retrieval pipelines. You own architecture decisions, you measure what you ship, and you reason about cost, failure, and quality before writing code. This is a broad, high-ownership role spanning automation infrastructure, applied ML, and knowledge/RAG systems. We are not looking for someone who has done all of it perfectly; we are looking for someone who has done enough of it in production to learn the rest fast.

About Us

We're building the AI system that automates the broker's real work end to end: logging into hostile third-party portals, reading messy regulated documents, scoring risk, and making recommendations, reliably enough to trust in a domain where mistakes have consequences. The hard problems here are the good kind: anti-bot evasion at scale, applied ML on real outcome data, and RAG that has to cite the right clause from the right document version, every time.

We're a startup, which means you'll own architecture, not tickets. A few others are chasing this space, we intend to outbuild them by shipping fast and shipping things that actually work. Leading that charge is our CTO, Zain Ali, who recently left Google to build this, after engineering at Netflix and Microsoft. You'll work directly under that bar.

If you want to build the thing that defines a category, not maintain something that already won, this is the room to be in.

Responsibilities

• The Work
  
  

Distributed Automation

• Rearchitect a single-machine automation engine into a system handling 1,000+ concurrent jobs against third-party portals: microservices, persistent browser session pools, container orchestration
• Design the job queue and execution model: dispatch, retry, per-target rate limiting, circuit breakers, real-time status feed to the user-facing UI
• Build an onboarding engine that, given a portal URL and credentials, discovers fields, maps them to our internal schema, and generates a working automation script with no manual scripting
  
  

ML & Model Quality

• Replace heuristic scoring with supervised models trained on historical outcome data; design and train domain-specific risk and recommendation models
• Benchmark against LLM-only baselines and pick the right tool per problem
• Stand up eval harnesses on every model, runnable on demand and on a schedule, that catch regressions before production
• Own the R&D track for self-hosted inference (LLMs, embeddings, ASR, TTS) and deliver cost-vs-quality memos that drive build/buy decisions
  
  

Hard Requirements

• Python (FastAPI, async/await): production service design, not scripts
• Distributed job execution with Celery + Redis or equivalent, Postgres as job store; you have debugged retry storms and backpressure in production
• Kubernetes or Docker Swarm: deployed and operated, not tutorial-level
• Browser automation at scale: Playwright or Puppeteer handling MFA flows, session persistence, and anti-bot systems (isTrusted events, PerimeterX-class protection)
• Python ML stack: PyTorch or equivalent, scikit-learn, HuggingFace; you build training pipelines, not just call APIs
• Supervised learning and embeddings: labels, class imbalance, and clean feature pipelines from messy source data
• Eval-first mindset: every model and knowledge system ships with a benchmark for correctness, freshness, and attribution
• You reason about cost-per-job, failure blast radius, and queue depth before writing code
  
  

Strong Plus

• LLM integration for DOM and field inference (vision models)
• LLM fine-tuning / RLHF
• SageMaker or equivalent for model serving; inference cost benchmarking across providers (Anthropic, OpenAI, self-hosted)
• Regulated-domain experience (insurance, fintech, health, legal) — knowing what you are reading matters
• Web scraping and change-detection infrastructure (PDF diffing, DOM monitoring)
• DigitalOcean infrastructure experience