Raza Jahangir

AI Systems Engineer · agents, RAG, and self-hosted AI infrastructure for production use.

Build → Evaluate → Deploy → Monitor

I work across the full lifecycle. That means designing the agent or retrieval pipeline, testing it against real data before release, running it as a reliable service, and tracking quality, cost, and latency once it's live.

Before AI, I spent 13+ years as a backend engineer building PHP applications, WordPress plugins, CodeIgniter apps, APIs, and databases, plus web scraping in Python. That background shapes how I build AI systems: clean data pipelines, stable integrations, and code that is easy to maintain after launch.

Focus

Agents — multi-step workflows with tool calling, state, memory, and human-in-the-loop steps (LangGraph)
RAG — ingestion, chunking, hybrid search, re-ranking, cited answers (LlamaIndex, pgvector)
Evaluation — test sets and regression checks before release (LangSmith)
Deployment — FastAPI services, self-hosted models, tracing, cost and latency tracking

Stack

AI / Python: Python LangGraph LangChain LlamaIndex LangSmith FastAPI LiteLLM Ollama
Data: PostgreSQL + pgvector Supabase MySQL
Web & infra: Node.js Next.js PHP WordPress Docker Cloudflare Tunnel

Projects

inhouse-ai · open source

A self-hosted, private AI stack for teams that want to run AI on their own hardware without sending data to third parties. Runs on anything from a CPU-only laptop to large GPU servers.

One Docker Compose file: Open WebUI, LiteLLM gateway, Ollama, PostgreSQL + pgvector, Cloudflare Tunnel
Document Q&A with citations; each user's knowledge bases are isolated
No inbound ports open; identity check before the login page
Python benchmark tool measuring time to first token and tokens/s on your own hardware
Upgrade path to vLLM on bigger GPUs by changing only the gateway config

DraftCraft · open source

WordPress plugin that generates scheduled blog drafts in a site's brand voice.

Bring-your-own-key through OpenRouter
Review queue: nothing publishes without approval
SEO sync with Rank Math, Yoast, and AIOSEO

Atlas Planner · client work

AI goal-planning platform for a coaching membership community (WordPress + BuddyBoss). Members take a life & business assessment, get AI-recommended focus areas and goals, and turn them into scheduled actions. MVP built in 8 weeks, then developed in weekly releases for seven more months.

AI layer

OpenAI ranks focus areas from assessment scores and proposes SMART goals and actions, retrieved from the client's goal and action library
Weekly check-in analysis that generates short nudges and reschedule suggestions
Every response follows a JSON schema, is validated in PHP before saving, and is retried if invalid; tool calling for goal selection
All calls go through a server-side proxy; prompts and schemas are versioned in the repo with unit tests
Cost controls: per-call token budgets, input truncation, retries with backoff on 429/500
Telemetry per call (model, tokens, latency, estimated cost) in an admin view
AI kill switch that falls back to rules-based suggestions; PII stripped before sending, plus a "delete my AI data" option

Platform

Recurring actions with RRULE, timezone and DST safe; due date and recurrence rule enforced as mutually exclusive at the API and database layers
Background jobs with Action Scheduler: hourly reminders, daily digests, weekly check-ins
REST API, custom database tables, Elementor widgets
Integrations: BuddyBoss notifications, ActiveCampaign email/SMS and CRM sync, GamiPress points and streaks
Exports to PDF, iCal, and CSV

Custom GPT Manager · client work

WordPress plugin for building and deploying multiple custom GPTs from wp-admin, each with its own knowledge base, access rules, and ChatGPT-style chat UI. Built end to end, then extended over six months.