Skip to content
TO

AI engineering

Agents that remember what matters.

Most of the work in a useful agent isn't the prompt. It's deciding what the agent should remember, for how long, and where that information lives, and then keeping it running in production.

Featured project

AI Sales & Support Agent

A production-oriented agent that integrates with an existing .NET business platform. Layered memory, RAG over company knowledge, tool orchestration and graceful degradation when parts fail.

  • FastAPI
  • Gemini 2.5
  • PostgreSQL
  • pgvector
  • Redis
  • Mem0
  • .NET
Read the case study

Agent core

FastAPI + Gemini 2.5

  • Working memory

    Redis

    What the agent is doing right now: active context, intent, state. Older turns roll up into summaries.

  • Knowledge memory

    pgvector + RAG

    What it knows: docs, FAQs and policies, retrieved with semantic search plus structured filtering.

  • User memory

    Mem0

    What it remembers about each user: preferences and history, so conversations continue.

Featured project · OpenClaw

Aria: a sandboxed agent for every employee

OpenClaw configured for Genuka: every employee has their own agent, sandboxed so memory never leaks between people. Group agents help manage each group, send reminders and track its goals. It runs across several LLMs, including GPT-4o and DeepSeek, on a Dockerised VPS. I wrote its custom skills and fixed the production issues that only appear with real traffic.

  • OpenClaw
  • GPT-4o
  • DeepSeek
  • Multi-LLM
  • WhatsApp
  • Docker
  • Coolify
Read the case study

Custom skills I built

  • BrainPre-call context manager: relevance filtering, history compression and token logging to cut spend.
  • Coding assistantPlanner → Developer → Reviewer pipeline with model routing, JSON contracts, a capped fix loop and a hard rule against pushing to main.
  • Humanize-writingChecks WhatsApp replies and outreach for AI tells before anything is sent.

  • MCP

    Exploring the Model Context Protocol: define tools once, reuse them across agents and models, keep the agent lean.

  • Agent tooling

    For Geniusland's lesson agent I built a safe editor for non-technical staff and a PDF-to-JSON ingestion pipeline.