Docerity

AI & LLM Engineering

AI systems that run in production, not a demo.

LLM routing, RAG, multi-modal pipelines, and conversational agents — processing millions of real requests, plus the model evaluation work behind training data that actually improves a model.

See the systemsStart a project
Systems
04
Models
07
Capabilities
04

In production

4 systems

Systems

Built for real traffic, not a proof of concept.

SmartLLMRouter

90% cost optimization

A 3-tier fallback system (GPT-4o → Gemini-2.0 → rule-based) processing millions of social media posts, routing each one by complexity instead of sending everything to the most expensive model.

  • GPT-4o
  • Gemini
  • Cost routing

Production RAG system

Real-time across 3 platforms

MongoDB Atlas Vector Search over 1536-dim embeddings, with hybrid search combining cosine similarity and metadata filtering — powers real-time intent analysis across Twitter, Facebook, and TikTok.

  • RAG
  • Vector search
  • MongoDB Atlas

Multi-modal vision pipeline

2,800+ LOC production service

GPT-4o Vision, DALL-E 3, and custom computer vision (blur detection, exposure analysis) behind quality gates and platform-specific optimization for LinkedIn, Instagram, Twitter, and Facebook.

  • GPT-4o Vision
  • DALL-E 3
  • Computer vision

WhatsApp AI assistant

9-container production deployment

A conversational assistant that turns a message or product photo into published social content — intent detection, image analysis, and multi-turn context memory, deployed across Nginx, FastAPI webhooks, and Celery workers for parallel processing.

  • Conversational AI
  • FastAPI
  • Celery

What I build

The part after the demo works.

Getting a model to answer once is the easy part. These are the pieces that keep it answering — cheaply, reliably, and at real volume.

  • LLM integration & routing

    01

    Multi-provider setups (GPT-4o, Claude, Gemini) with complexity-based routing so cost scales with actual difficulty, not worst-case.

  • RAG & vector search

    02

    Production retrieval systems with real embeddings and hybrid search, not a demo that falls over past a few hundred documents.

  • Multi-modal pipelines

    03

    Vision, image generation, and text working together behind real quality gates, tuned per platform.

  • Model evaluation & RLHF

    04

    Designed training samples and evaluation rubrics, and annotated model responses across 150+ technical tasks — I know what makes training data actually improve a model.

Models & platforms

Picked per task, not one model for everything.

Each call goes to the model that fits it, weighing quality against cost instead of paying top price for every request.

In the toolbox

  • GPT-4o
  • Claude 3.5
  • Gemini
  • DALL-E 3
  • Stability AI
  • OpenAI Assistants API
  • MongoDB Atlas Vector Search

Let's talk

Got something worth building?

A project, mentorship, or just a question — I read every email myself.

Start the conversation
  • Replies in 1–2 business days
  • Goes straight to me

docerity — start