Skip to content

Built & written by Luca Viscomi

Engineering AI.
Sharing the build.

Luca Viscomi

AI Engineer · Founder · Builder

Production AI agent engineering, Model Context Protocol tooling, multi-agent architecture and context optimisation — built, tested and documented in public.

MSc Artificial Intelligence · Distinction

⚡ Get MCP Harness Explore the builds View my profile
  • Python
  • MCP
  • LangGraph
  • AWS
  • Azure
  • GCP
Luca Viscomi in a navy suit, surrounded by electric blue orbital light rings

Explainer Series

Get Your Head Around Agents

Plain-English guides for managers, executives, and leaders moving from AI hype to practical business execution.

VIDEO BRIEFING · 1:25
0:00 / 1:25
EXPLAINER · PART 01✨ Narrated Video Intro

Chatbot vs Copilot vs Agent

The plain-English executive guide to what AI actually does in business.

Watch the 1:25 video intro on the left for the executive breakdown of tools that talk vs systems that work, or read the full article.

Level 1 (Chatbots): Conversational research assistants with 0% state execution
Level 2 (Copilots): Passenger-seat accelerators requiring human hands on the wheel
Level 3 (AI Agents): Digital operators taking custody from goal planning to verified execution
Read Full Explainer

Developer Tooling · Open Source

Agent Junky MCP Server

The AI Systems Engineering Harness for Claude Desktop & Cursor. Real token accounting, tool audits, and Layer 1–7 blueprints.

MCP 2.X READY · STDIO & SSE

4 Production Tools In Your Chat

Audit tool schemas, count real BPE tokens, harden system prompts, and retrieve LangGraph blueprints directly in Claude Desktop.

⚡ Setup in 30 Seconds
DEV TOOLS · HARNESS⚡ Local & Cloud

AI Systems Engineering Harness for Claude

Stop blaming the model when agents break in staging. Give Claude deterministic tools.

Connects directly into Claude Desktop or Cursor over standard I/O (stdio) or HTTP/SSE on Render.

audit_tool_architecture: Catches parameter-explosion bloat and ambiguous keys
benchmark_context_bloat: Real BPE token calculations, SNR, and cost estimates
retrieve_system_blueprint: Instant production code for LangGraph supervisors & SQLite memory
Explore MCP Server & Quickstart View All 4 Tools

Telemetry · Active Optimization

Tokenomics & Context Telemetry

Live session telemetry, deterministic transcript surgery, and repository token budgeting to defeat runaway context bloat.

VIDEO BRIEFING · 1:24
0:00 / 1:24
ENGINEERING HARNESS · v1.2.x⚡ Active Telemetry

Tokenomics & Context Bloat Optimization

Stop burning tokens on zombie terminal logs, test failures, and unchecked context saturation.

Watch the 1:24 video intro for the complete breakdown of transcript surgery, or launch the interactive browser simulator.

Active Transcript Surgery: Prunes dead output logs into semantic tombstones with 60–80% savings
Repository Token Auditor: Scans multi-file codebases and projects cold vs warm invocation costs
Session State Handover: Synthesizes zero-loss continuation files for effortless context window resets
⚡ Launch Tokenomics Engine 📘 Beginner's Guide & KB

Conversational Telephony · 2026 Benchmark

AI Voice & Call Agents

Sub-600ms latency pipelines, interactive ROI telemetry, and in-depth engineering comparison across the 8 leading platforms.

Open Comparison Engine
Autonomous AI Voice Assistant Telephony Infrastructure and Call Agents 2026
BENCHMARK • SUB-600MS LATENCY
CONVERSATIONAL TELEPHONY • 2026 BENCHMARK🎙️ Interactive Engine Live

The 8 Best AI Call & Voice Assistants in 2026

Sub-600ms latency pipelines, turn-taking & barge-in, HIPAA guardrails, and real-time telephony pricing.

Comprehensive engineering comparison of Bland AI, ElevenLabs, Retell, Vapi, Synthflow, Air AI, Phonely, and PlayAI across enterprise latency, SIP trunking, and compliance.

Latency Benchmarks: Sub-600ms voice round-trips & barge-in evaluated across 8 leading platforms
Interactive ROI Calculator: Dynamic telephony minutes vs receptionist payroll ($40k–$80k+ annual savings)
Compliance & Guardrails: HIPAA BAA agreements, audio PII masking, and SIP warm-transfer failovers
🎙️ Launch Voice AI Platform & Calculator 📖 Read In-Depth Guide

AI Infrastructure · Interactive Lab

The AI Storage Atlas

Disaggregated NVMe fabrics, MLPerf v3.0 checkpoint telemetry, and sovereign cloud reference architectures.

Open Storage Atlas
The AI Storage Atlas Distributed Storage Architecture Lab 2026
INTERACTIVE LAB • MLPERF STORAGE V3.0
AI INFRASTRUCTURE • 2026 ARCHITECTURE⚡ Interactive Lab Live

The AI Storage Atlas: From SAN to Sovereign AI

Disaggregated NVMe fabrics, VAST DASE, Ceph, WEKA, and sovereign AI cloud data platforms.

Explore interactive architecture diagrams, calculate GPU idle-time financial stalls, compare erasure coding math, and inspect 1.25T parameter checkpoint benchmarks.

Interactive Architecture Lab: Toggle between Legacy SAN, Ceph, VAST DASE, WEKA, DDN, and NetApp AFX
GPU Waiting-Time Economics: Model cluster GPU-hour waste ($897k+/yr on 256 H100 clusters) from I/O starvation
Sovereignty & MLPerf Benchmarks: Core42/G42 reference design and 877 GiB/s 1.25T checkpoint telemetry
⚡ Launch AI Storage Atlas 📖 Read In-Depth Guide

GPU Processing Pipelines · Deep Dive #18

AI Training vs Inference on GPUs

Inside the GPU processing pipeline: backpropagation, distributed all-reduce, autoregressive decoding, and KV-cache architectures.

Read Full Deep Dive
AI Training vs Inference on GPUs: Processing Pipelines, Memory Hierarchies, and Data Flow
DEEP DIVE #18 • PIPELINE SIMULATOR
DEEP DIVE · #18 / GPU PROCESSING PIPELINES⚡ Latest Deep Dive

AI Training vs Inference: Inside the GPU Pipeline

From backprop and all-reduce synchronization to token generation, KV caching, and memory offloading.

The architectural breakdown of GPU compute, memory consumption, storage dependencies, and networking topologies across AI model training and LLM inference. Includes interactive 7-stage vs 5-stage pipeline simulation.

Compute Demands: Training requires 3× FLOPs/token (1× forward + 2× backward); inference executes compute-dense prefill then memory-bandwidth-starved decode
Memory Physics: Training locks 16 bytes/parameter in FP32 Adam states + activations; inference locks model weights + dynamic O(N) KV cache
Interconnect Topologies: Lossless 400G–800G InfiniBand all-reduce synchronization vs disaggregated North-South API request routing
⚡ Read Deep Dive #18 ⚙️ Interactive Simulator

GPU Computing · Inference Memory Architecture

Inside the AI Memory Wall

GPU HBM3e, dynamic KV-cache scaling, GPUDirect Storage, and WEKA NeuralMesh distributed memory fabrics.

Read Full Deep Dive
Inside the AI Memory Wall: GPU HBM, KV Cache, NVMe Fabrics and WEKA NeuralMesh Architecture
DEEP DIVE #17 • INTERACTIVE CALCULATOR
DEEP DIVE · #17 / AI MEMORY HIERARCHY⚡ New Release

Inside the AI Memory Wall: GPU HBM, KV Cache & WEKA

Why 70B models starve 8-way GPU nodes during 128K multi-agent inference—and how to fix it.

How modern inference engines turn GPU HBM, host DDR5, local NVMe, and distributed parallel storage into a unified memory tier. Includes interactive KV-cache capacity modeling and GPUDirect data paths.

The Inference Dilemma: Model weights are static, but KV cache scales with context (32x 128K streams = 524 GB)
WEKA Augmented Memory Grid: Microsecond-scale block restoration directly into GPU interposers via GPUDirect RDMA
Interactive Capacity Calculator: Real-time HBM vs offload projection across Llama-3-70B, Mixtral, and DeepSeek
⚡ Read Deep Dive #17 🧮 Interactive Calculator

Architectural Taxonomy · Topic Hubs

Engineering Topic Clusters

Structured knowledge hubs grouping our foundational pillars, protocol implementations, and benchmarks.

Explore All Clusters
🤖 AI Agent Architecture (10 Guides)🔌 Model Context Protocol (4 Guides)🧠 Context Engineering (7 Guides)⚡ AI Token Optimisation (3 Guides)🛡️ Agent Reliability (7 Guides)

Engineering Build Log

Deep Dive Series

In-depth technical blueprints, runtime harnesses, MCP protocols, and multi-agent systems architectures.

View All 18 Guides Follow via RSS

DEEP DIVE · #01 / THE LOCAL LAB

Build a home agent lab

Windows, Python and Google ADK. A real local build, mistakes included.

DEEP DIVE · #02 / AGENT ARCHITECTURE

Speed. Control. Governance.

Four agent-building paths. The trade-offs behind choosing the right one.

DEEP DIVE · #03 / MCP & TOOLING

Giving the agent hands

Wiring local MCP servers with Python and SQLite. Connect your tools. Keep control.

DEEP DIVE · #04 / TOOL ARCHITECTURE

Engineering Reliable AI Agents

Tool routing, least-privilege scoping, and evaluation. Avoid the tool-explosion trap.

DEEP DIVE · #06 / TOKENOMICS

Deterministic BPE Token Accounting

Why Byte-Pair Encodings break multi-agent loops and how to fix subword fragmentation.

DEEP DIVE · #07 / MCP GATEWAYS

Production MCP: Streamable HTTP

Streamable HTTP architecture, high-throughput agent gateways, and legacy SSE migration.

DEEP DIVE · #10 / SESSION SURGERY

The O(N²) History Re-feeding Trap

How multi-turn transcripts explode token bills to 2.6M tokens, and surgical pruning fixes.

DEEP DIVE · #13 / STATE PERSISTENCE

Zero-Loss Context Resets

Maintaining long-running agent coherence with structured state handover checkpoints.

DEEP DIVE · #15 / VOICE TELEPHONY

The 8 Best AI Call & Voice Assistants

Bland AI, ElevenLabs, Retell, Vapi & 4 others benchmarked on latency, HIPAA security, and ROI.

DEEP DIVE · #16 / STORAGE ARCHITECTURE

From SAN to Sovereign AI Storage

NVMe fabrics, VAST DASE, Ceph, WEKA, DDN, and sovereign AI cloud data platforms in 2026.

DEEP DIVE · #17 / AI MEMORY HIERARCHY

Inside the AI Memory Wall

GPU HBM, KV cache scaling, GPUDirect Storage, and WEKA NeuralMesh distributed memory fabrics.

DEEP DIVE · #18 / GPU PROCESSING PIPELINES

NEW

AI Training vs Inference on GPUs

From backprop and all-reduce to token generation, KV caching, and memory offloading.

VIDEO BRIEFING · 2:35
0:00 / 2:35
DEEP DIVE · #05 / SYSTEMS ARCHITECTURE✨ Narrated Video Intro

7 Layers of AI Systems Engineering

From foundation models to runtime harnesses, agent loops, and enterprise ontologies.

Watch the narrated video intro on the left for the rapid bird's-eye view, or dive directly into the complete production engineering blueprint.

Tiers 1–3: Foundation Models, Prompt Schemas & Context Boundaries
Tiers 4–5: Execution Harnesses & Autonomous LangGraph Loops
Tiers 6–7: Multi-Agent Orchestration & Domain Ontologies
Read Full In-Depth Article
📚 Explore All 18 Technical Articles & Blueprints