Work

One forward deployed engagement, written up as a case study, then other projects across agentic AI, RAG, document AI and voice.

Case study

Air-gapped bilingual AI assistant for a government client

A bilingual English/Arabic generative AI assistant, deployed on a government client's on-premise, air-gapped servers. Working in a team of three, I took on most of the technical delivery, from discovery sessions and GPU benchmarks to the on-site release deployments and the handover for user acceptance testing.

  1. Discovery
  2. Requirements
  3. Design
  4. Build
  5. Air-gap
  6. Deploy
  7. UAT handover

Deployed on-premise and handed over for user acceptance testing in October 2026.

Read the case study

More projects

  • On-Prem Enterprise AI Platform

    Self-hostable enterprise AI platform orchestrating LLM tool-calling across a federation of MCP servers, backed by a fully self-hosted inference stack.

    Stack: MCP, vLLM, Qdrant, Keycloak, Docling

  • Hybrid RAG Sales Intelligence

    Proposal search assistant fusing vector search, graph traversal, and Reciprocal Rank Fusion behind a layered security pipeline.

    Stack: Qdrant, Neo4j, Llama Guard, WebSocket, Celery

  • Regulatory Document Verification

    RegTech assistant verifying governance reports against regulatory articles with confidence-based LLM validation.

    Stack: LangGraph, Qdrant, OpenRouter, FastAPI

  • Vision-Based Document Extraction

    Multi-format document-to-JSON pipeline combining OCR, vision-language models, and local OCR inference.

    Stack: Docling, GPT-4 Vision, Gemini, DeepSeek-OCR, FastAPI

  • AI-Powered Log Analytics Platform

    Enterprise log analytics with multi-agent orchestration, semantic error grouping, and anomaly detection.

    Stack: Multi-Agent, Qdrant, Redis Pub/Sub, Langfuse, Presidio

  • AI-Powered Outreach Automation

    Agentic LLM platform orchestrating multi-phase target discovery, data enrichment, fitness scoring, and branded document generation via MCP server architecture.

    Stack: Claude, MCP, Pydantic, python-pptx, openpyxl

  • Agentic AI for Roadside Assistance

    Conversational AI platform with real-time voice synthesis, sentiment analysis, and intelligent technician routing.

    Stack: ElevenLabs, GPT-4o, Claude, WebSocket

  • AI-Powered Debt Collection

    Voice AI platform with multi-model analysis pipeline for automated loan recovery conversations.

    Stack: ElevenLabs, Silero VAD, ONNX, WebRTC

  • Production RAG System

    Semantic search with two-stage retrieval, cross-encoder reranking, and content guardrails for specialized domains.

    Stack: Sentence-Transformers, Qdrant, Cross-Encoders, FastAPI

Search and jump

Home Start page
Work All projects
Air-gapped bilingual AI assistant for a government client Case study
About How I work
Blog All posts
Context and constraints Case study section
My role Case study section
Discovery to requirements Case study section
Architecture Case study section
Key decisions Case study section
Deployment and rollout Case study section
Making Arabic PDF text searchable Case study section
Bringing a server clock back in sync Case study section
Re-tuning the Arabic model Case study section
Handover Case study section
Results Case study section
What Iā€™d carry into the next engagement Case study section
How I worked Case study section
How I load-test self-hosted model serving before an air-gapped rollout November 3, 2026
I Stopped Using Basic RAG in 2026. Here's What Replaced It. March 16, 2026
Building a Multi-Agent Log Analytics Platform From Scratch March 14, 2026
The Secret Life of a Data Detective January 24, 2026
Why You Don't Use a Sledgehammer to Crack a Nut January 21, 2026
Building an Intelligent Document Extraction System with Vision-Language Models January 2, 2026
Building Enterprise Agentic AI Systems: Lessons from Production Voice Agents January 1, 2026
Open resume PDF, new tab
Send an email arnolps172@gmail.com
GitHub github.com/Arnol-P-S
LinkedIn linkedin.com/in/arnol-ps
Switch theme Light or dark