LabVoice & Multimodal AI
Arabic & Darija Voice Agent
Speech-to-speech for Moroccan Darija: dialect-tuned ASR, intent routing with a human-handoff threshold, and local TTS inside a conversational latency budget.
WhisperSparkTTSPyTorchLoRA
The lab
Things I build to understand them.
Smaller than the production work and labelled as what it is. Protocol implementations, retrieval pipelines, training runs and the field notes behind them.
Experiments
7
LabAgentic Systems & Orchestration
MCP Server & Client Toolkit
A Model Context Protocol server exposing real tools, with a matching client and a test suite — connected to a desktop LLM host.
PythonMCPpytestuv
LabAgentic Systems & Orchestration
A2A Orchestrator & Worker
Agent-to-Agent protocol demo: an orchestrator discovering a worker through its agent card and delegating a task across the wire.
PythonA2AAgent cardsHTTP
LabAgentic Systems & Orchestration
Agent Patterns Catalog
Five runnable agent patterns with diagrams: prompt chaining, routing, parallelization, the agent loop, and structured output.
PythonLLM APIsPydantic
ConceptRAG, Graph RAG & Retrieval
Agentic RAG Service
Retrieval service with query rewriting, reranking, self-correction and an eval harness that scores retrieval separately from generation.
FastAPIQdrantLangGraphRagas
LabFine-tuning & Alignment
LoRA → DPO → GRPO Pipeline
An end-to-end alignment pipeline: data prep, SFT with TRL, evaluation, then preference optimization with DPO and GRPO.
PyTorchTRLPEFTHF Datasets
ConceptLocal LLM, GPU & Inference
LLM Cost & Inference Dashboard
Local-vs-API routing, semantic cache hit rate, GPU memory headroom, and the batch-size / latency trade-off made visible.
vLLMOllamaRedisPrometheusGrafana
Field notes
14Field notes on building production AI systems — collected, verified and synthesized from real deployments and post-mortems. These are lessons and patterns, not project write-ups.
01Agentic Systems & Orchestration8 recurring failure modes
AI Agents in Production — What Actually Breaks
The eight failure modes that repeat in every agent system, and the fix for each.
Read the note →
02Agentic Systems & Orchestration36% error accumulation by step 20
Multi-Agent Orchestration — When to Use It and When to Avoid It
80–90% of production scenarios succeed with an orchestrated workflow. Most teams reach for multi-agent anyway.
Read the note →
03Agentic Systems & OrchestrationTracing on every tool call
Why 90% of Demo Agents Fail in Production
The model is the easy part. Observability, error handling and evaluation are where agents live or die.
Read the note →
04Local LLM, GPU & Inference88% cost reduction
From $18K to $2.1K/Month on a Multi-Agent System
Cutting 88% of the bill by bringing the model in-house — and fine-tuning it to actually use tools.
Read the note →
05Agentic Systems & Orchestration$847 in one night
The Agent Loop of Death
One missing termination condition, six hours of two agents talking to each other, and an $847 invoice.
Read the note →
06Local LLM, GPU & Inference14s → 4s
Cutting Multi-Agent Latency from 14s to 4s
Stop using an LLM for tasks that are pattern matching. Measure every step before optimizing any of them.
Read the note →
07Local LLM, GPU & Inference91% cost reduction
From $52K to $4.8K/Month on a Customer Service AI
High API cost and data privacy look like two problems. They are usually the same problem with the same fix.
Read the note →
08AI Backend Engineering20–30 patterns cover 80%
GenAI System Design — What Nobody Tells You
20–30 query patterns cover 80% of use cases. Context management, not model choice, is the real bottleneck.
Read the note →
09Fine-tuning & Alignment2K clean > 10K noisy
Fine-Tuning — When It Makes Sense and When It Doesn't
A six-question decision framework, and why 2,000 clean examples beat 10,000 noisy ones.
Read the note →
10Local LLM, GPU & Inference150ms → 450ms
LLM Inference — The Batch Size Trap
Raising batch size from 8 to 32 gained 40% throughput and tripled latency. LLM inference is memory-bound, not compute-bound.
Read the note →
11Agentic Systems & OrchestrationProtocol vs. behaviour
MCP in Production — What Works and What Breaks
The protocol is solid. LLM behaviour around it is the bottleneck.
Read the note →
12Security & CompliancePrompts are not permissions
The Model Is Not a Trust Boundary
Everything an LLM sees is input and everything it emits is untrusted. Authorization has to live in code the model cannot reach.
Read the note →
13Voice & Multimodal AIWER 18% → 7%
Building a Voice AI Agent for Arabic Call Centers
Transcribe and categorize 500 calls before writing a line of code. The distribution tells you your ROI ceiling.
Read the note →
14Voice & Multimodal AI88% resolved
Building a Production Voice Agent — Lessons Learned
Resolving 88% of calls without a human. What actually took the time, and why it wasn't the model.
Read the note →The systems these lessons came out of are written up in full.
See the production work →
MAJID
“Innovation distinguishes between a leader and a follower.”
Need an AI system that holds up in production? Tell me what breaks today and I'll tell you what I'd build.