Internship Diary Entry Details Internship: AI Engineer - SynerSense

Date: 14 Mar 2026


Deployment Goal

Deploy the Relabeling Module and Local RAG Vault infrastructure onto the client laptop, transitioning the system from development staging into a stable, privacy-first local setup.


What I Worked On

Infrastructure Setup (Local + Cloud Bridge)

  • Configured a PostgreSQL + pgvector container via Docker.
  • Configured persistent volumes so the database state survives container restarts and OS reboots.
  • Set up a secure connection between the cloud-hosted SvelteKit frontend and the local FastAPI backend via Cloudflare Tunnel, avoiding direct exposure of local ports.

Workspace Persistence (Reference State)

  • Verified the .state/ directory and initialized references.json.
  • Ensured selected reference thumbnails are stored locally and restored automatically on reload.

Local Embedding Workflow (Performance & Stability)

  • Loaded the Sentence Transformers embedding model on the client CPU.
  • Measured and optimized response latency to achieve acceptable local similarity-query performance.

Smoke Testing (UX + Functionality)

  • Validated Reference Selection Mode and Relabel Mode end-to-end.
  • Confirmed reference thumbnails render correctly inside the scatter plot.
  • Verified hover previews and UI responsiveness under annotation workloads.

Hours Worked

8


  • Docker - https://docs.docker.com
  • pgvector - https://github.com/pgvector/pgvector
  • Cloudflare Tunnel - https://developers.cloudflare.com/cloudflare-one/connections/connect-networks/
  • FastAPI - https://fastapi.tiangolo.com
  • SvelteKit - https://kit.svelte.dev/docs
  • Sentence Transformers - https://www.sbert.net

Learnings & Risks

Key Learnings

  • On-site deployment of AI systems on client hardware while preserving compatibility with existing infrastructure.
  • Working with Docker volume mappings between Windows hosts and Linux containers.
  • Designing privacy-first AI architectures where vector databases and embeddings run locally.
  • Communicating vector search and embedding concepts to non-technical stakeholders.

Blockers / Risks

  • The client’s laptop has lower memory capacity than the development machine, so monitoring embedding/model usage is required to keep the UI responsive.
  • The current JSON-based workspace state works, but users need a reset mechanism to quickly clear their reference workspace.

Skills Used

Python, FastAPI, SvelteKit, Docker, PostgreSQL, pgvector, AI Deployment, System Configuration, Local AI Infrastructure, Debugging, Client Support, Documentation.


This site uses Just the Docs, a documentation theme for Jekyll.