AI Vision Processors For Edge Applications

Our solutions make cameras smarter by extracting valuable data from high-resolution video streams.

Job Description

Key Responsibilities

1. Design & Implement Complex Agentic Workflows

• Architect an end‑to‑end inspection pipeline that:
  • Perceives: Fuses multi‑modal inputs (high‑res images, thermal, acoustic, IMU) in real time.
  • Reasons: Applies an RL‑trained policy to triage anomalies into severity categories (e.g., Critical, Warning, Normal).
  • Acts: Generates a structured maintenance recommendation (e.g., “Replace bearing A; confidence 92 %”) and logs the full decision path.
• Design the workflow to be modular, testable, and reusable — serving as a template for future sustainability agents.

2. Train & Optimize RL‑Based Inspection Policy

• Use policy optimization (e.g., PPO, A3C, or custom reward‑shaped algorithms) to train the agent on custom synthetic + real‑world datasets you curate.
• Define dense reward functions that reflect sustainability KPIs: energy waste, emission impact, component wear, and safety risk.
• Apply reward shaping to avoid local optima and ensure the policy generalizes across facility types (e.g., manufacturing lines, HVAC systems, solar farms).

3. Build High‑Quality Custom Datasets

• Curate, label, and augment a dataset of ≥ 10 k annotated samples covering diverse failure modes (e.g., corrosion, mis‑alignment, thermal leakage).
• Implement data‑quality guardrails (e.g., label consistency checks, sensor calibration validation) to guarantee the RL policy learns from trustworthy signals.
• Publish dataset documentation and preprocessing pipelines so the broader team can reuse your standards.

4. Embedded‑First Optimization

• Port the trained policy to Ambarella’s edge‑AI dev kit.
• Achieve real‑time inference (≤ 50 ms latency per decision) via:
  • Model distillation / quantization for Ambarella’s AI accelerators.
  • Memory‑efficient tensor layouts to keep RAM footprint < 128 MB.
  • Thread‑safe, lock‑free execution to meet deterministic deadlines.
• Profile and report per‑module latency and energy per inference; optimize for ≤ 0.5 W average power during continuous operation.

5. Audit‑Trail & Explainability

• Embed a self‑logging audit trail that captures, for every inspection:
  • Why the agent flagged the input (e.g., “Thermal hotspot ΔT = 38 °C > threshold, located at coordinate x=120,y=45”).
  • What it recommends (structured JSON: { "action": "schedule_maintenance", "component": "pump_bearing_3", "priority": "high", "confidence": 0.92 }).
  • How confident it is (calibrated probability score + entropy of the policy).
• Ensure the trail is human‑readable, tamper‑evident, and exportable in JSON/CSV for compliance audits.

6. Validation & Scenario Testing

• Execute a scenario suite covering edge cases:
  • Sensor degradation (e.g., partial lens obscuration, thermal drift).
  • Ambiguous severity (e.g., early‑stage corrosion vs. surface stain).
  • High‑noise environments (vibration, intermittent connectivity).
• Achieve ≥ 90 % precision in severity classification and ≤ 5 % false‑negative rate across all scenarios.

Required Qualifications

• Pursuing B.S., M.S., or Ph.D. in Computer Science, AI/ML, Robotics, or related field (graduating 2026‑2027).
• Strong programming skills: Python (PyTorch/TensorFlow, RLlib/Stable Baselines3), C/C++ for embedded integration.
• Demonstrated experience with:
  • Designing multi‑stage agentic workflows (perception → reasoning → action).
  • Reinforcement learning: policy optimization, reward shaping, or custom RL algorithms.
  • Building or curating large, high‑quality datasets (image, sensor, or multimodal).
• Familiarity with embedded Linux, performance profiling (e.g., perf, vtune), and low‑power optimization.

Preferred Skills

• Experience with sensor fusion (thermal + RGB + IMU) or industrial IoT protocols (Modbus, OPC-UA).
• Knowledge of sustainability metrics (CO₂ impact, OEE, MTBF) and how to embed them in reward functions.
• Use of behavior‑tree or rule‑based fallbacks to guarantee safety when the RL policy is uncertain.
• Prior work with edge‑AI accelerators (NPU, DSP, or Ambarella’s AI Engine).