Internship Diary Entry

Role: AI Engineer - SynerSense Date: 17 February 2026 Working Hours: 09:30 AM - 06:30 PM Total Hours: 9

---

Work Summary

Production Inference Architecture Design - AnanaCare Relabel Platform

• Finalized migration to a single Hugging Face Inference Endpoint integrating:
  • Custom CLIP backbone (with intermediate layer embeddings)
  • Trained DNN model (dnnV1_model_best.pth)
• DNN Model Review:
  • Hidden layers: [512, 256, 128]
  • Attention: enabled
  • Early stopping: epoch 23
  • Best validation loss: 0.06306
  • Action: Verify embedding dimensional compatibility between CLIP outputs and DNN input (in_features of first linear layer)
• CLIP Embedding Extraction:
  • Ensured inference mirrors training-time behavior:
    • Correct transformer layer selection
    • Pooling strategy (CLS vs mean)
    • Normalization steps
    • Image preprocessing (resolution consistency)

• Final Inference Pipeline:

  Input Image →
  > Custom CLIP Forward (output_hidden_states=True) →
  > Intermediate Layer Extraction →
  > Pooling & Normalization →
  > Bias Feature Concatenation →
  > DNN Forward Pass →
  > Prediction Output

• Integration Risks & Mitigations:
  • Embedding mismatch
  • Missing categorical bias injection
  • Normalization inconsistencies
  • Preprocessing drift
  • Mitigation steps defined for each risk

---

Key Learnings & Outcomes

• Deepened understanding of production inference orchestration
• Importance of strict training–inference pipeline consistency to prevent distribution drift
• Embedding dimensional compatibility as a critical deployment validation step
• Strengthened architectural thinking (latency optimization, API dependency reduction)
• Shifted focus from UI-level debugging to infrastructure-level system design

---

Blockers & Risks

• Confirm DNN first layer input dimension for embedding size compatibility
• Retrieve exact training-time embedding extraction & normalization logic
• Deployment depends on precise preprocessing replication

---

Skills Utilized

• PyTorch model architecture analysis
• Transformer embedding extraction concepts
• Production inference design
• Risk assessment & mitigation planning
• System architecture thinking