2025 – present
VynixAI · Senior Solutions Architect, AI & Cloud Infrastructure
Building AI-driven, cloud-native platforms for real-time retail and enterprise workloads: RAG pipelines, scalable inference, and agent automation.
Open to consulting, collaboration, and roles
Free tools and plain-English notes for people building with AI.
This headline started as a wordy prompt. Cutting the filler took it from about 23 tokens to 12.
I’m Iggy Mwangi, an AI systems architect in Dallas–Fort Worth. Here you’ll find free tools that run in your browser, short essays on keeping AI systems working in production, and my open-source projects. You don’t need an account for any of it.
Now VynixAI · Nubla AI Earlier Apple · Contran · AT&T · Sam’s Club
What brings you here?
- I want to tidy up a prompt Shorten it, reuse it across chats, or compare two versions. Free, in your browser.
- I want to learn how AI systems hold up Essays and reference guides on context, cost, reliability, and incidents.
- I’m looking for code I can use Open-source guardrails, audit tools, and context utilities on GitHub.
- I’d like to hire or work with Iggy What I help with, my background, and how to get in touch.
Try it: make a prompt clearer
Paste a prompt, or pick an example. You’ll get a score, a list of what’s missing, a version with the filler words removed, and a rewrite with clear sections. It’s a small version of Prompt Matrix Evaluator, and everything happens in your browser.
Tightened
Hi! Write a welcome email for our new customers. Mention the free trial and stuff, and keep it friendly and good.
Tools
Each tool is a single page you can use right away. Every card says where your text goes, so you can choose the ones that fit your privacy needs.
New to this? Four words you’ll see a lot
- Prompt
- The instructions and text you give an AI model, like a message in a chat.
- Token
- The unit models read and bill by. A token is roughly three-quarters of an English word, so fewer tokens usually means faster and cheaper replies.
- Context window
- How much text a model can consider at once. When a conversation outgrows it, older parts get dropped or ignored.
- Agent
- An AI model that works in steps: it uses tools, checks the results, and tries again if something fails, instead of answering once.
What the labels mean
- Stays on your device Your text is processed in this browser tab and not sent anywhere.
- Remembers your work Saves to this browser’s storage so it’s there next visit. Clear your site data to remove it.
- Needs your API key Sends your prompt to OpenAI or Anthropic using a key you supply.
- Sends text to a server Your text goes to a server I run, which passes it to a hosted model.
Improve a prompt
-
Prompt Matrix Evaluator
Checks a prompt for a clear task, output format, length limit, audience, vague words, and contradictions, then cuts the filler and shows what that saves per month at your volume. A good first tool to try.
-
Context Extractor
Paste a long chat or notes and get the goals, rules, decisions, open questions, and key facts. Untick or re-sort any item, then copy a ready-made handoff prompt for a fresh chat.
-
Claude Context Engine
Rewrites messy logs, code, or notes into a tighter prompt using a hosted model. Secrets, emails, and phone numbers are masked before anything is sent, and you can preview exactly what goes out.
Build, reuse, and compare prompts
-
Portable Context Engine
Reusable instruction profiles with fill-in variables and defaults, six starter templates, and Markdown, XML, or plain-text output for any AI chat or system prompt.
-
Prompt Diff & Stager
Compare two prompt versions inline or side by side, fill in placeholders, and keep snapshots you can export, import, and roll back to.
-
NanoBanana Compiler
Describe components and connections to get a live architecture diagram. Download it as SVG, paste Mermaid into your docs, or copy a prompt for an image model.
-
Multi-Model Workspace
Ask OpenAI and Anthropic models the same question side by side, with follow-ups, a shared system prompt, response times, token counts, and a Markdown export.
Plan and design AI systems
-
Context Window Optimizer
Plan how much text fits in one request. Budget each part of the prompt, check there’s room for the reply, see the monthly cost, and find what to cut first.
-
Agent Stateboard
Sketch an AI agent as a fixed list of steps, each with a check and a limit on retries. Simulate failures, then export a spec, a diagram, or starter Python code.
-
Token Router
Enter your API rate limits and traffic to see whether you’ll hit them. Also see how urgent requests stay fast when things are busy, and how long a backlog takes to clear.
Writing
Two essays on this site, plus longer guides kept on GitHub.
- The shift from coding to context orchestration As more software runs on models, the leverage moves into how prompts are structured, bounded, and checked.
- Token window economics How long chat histories, multi-agent overhead, and strict output formats eat into response time.
- The AI Runbook What to do when an AI feature breaks: symptom, first action, diagnosis, fix, prevention. Written to be read mid-incident.
- Production AI Patterns Named patterns for the decisions that decide whether an AI system survives real traffic: context, cost, reliability, observability, and security.
- Agentic Engineering How to build AI agents that do real work: planning their steps, giving them the right context, and letting them use tools.
- Awesome AI Architecture A map of AI infrastructure tools, organized by the problem each one solves.
- Everyday AI, done right A companion site on using AI well in everyday work.
Projects
Open-source code on GitHub. Free to read, use, and adapt under each repo’s license.
- SurfaceLock Records the prompts, models, and settings your code depends on, so changes to them show up in review.
- ModelBump A behavioral diff for model upgrades and deprecations.
- AIBoM Generates an audit-ready AI Bill of Materials from your repo in one command.
- Deterministic Agent Supervisor The model suggests actions, but a fixed set of rules decides what’s allowed. Anything not on the list is blocked.
- Factory Gate Checks an AI agent’s code changes against your rules before they reach your main branch.
- Sliding Window Trim Engine Removes repeated parts of a long chat before it’s sent to the model, so each request costs less.
- CircuitX Research: finds which internal features a code model uses when it writes code.
- Utility Belt Small everyday browser utilities. No AI required.
Seventeen sites, one per subject
Pattern libraries, an incident runbook, field notes, a weekly briefing, a bench of small tools, and a few essays about maps, minds, and reading.
-
Flagship · synthesis
Applied AI Engineering
The hub. Start here for the big picture: it connects the ideas from the other sites into one argument about where AI systems are heading.
-
Reference · pattern language
Production AI Patterns
-
Reference · incident procedures
The AI Runbook
-
Guide · agent systems
Agentic Engineering
-
Workbench · 24 tools
Systems Bench
Work with me
Open to consulting, collaboration, and roles.
What I help teams with
- Making AI agents dependable. Fixed workflows and structured outputs, so an agent follows a path you can audit instead of improvising.
- Keeping costs and wait times down. Tighter prompts and smarter context so the important instructions stay in view and bills stay predictable.
- Shipping safely. Checks on tool calls and code changes before they land, with isolated execution and clear policies.
Recent work
2024 – present
Nubla AI · Cloud Architect and Machine Learning Engineer
ML systems and cloud architectures across AWS, Azure, and GCP, focused on LLM integration, retrieval, and dependable ML pipelines.
Results
- 38%
- fewer hallucinations after re-architecting our RAG pipeline VynixAI
- 42%
- more reliable multi-tool routing, by treating tools as state transitions VynixAI
- 27%
- lower spend from cost-aware LLM routing, with no drop in quality VynixAI
- 70%
- less debugging time with OpenTelemetry reasoning-trace observability VynixAI
- 22%
- lower LLM latency by reworking tokenization paths Nubla AI
- 35%
- shorter training time by reorganizing data across S3, BigQuery, and ADLS Nubla AI
Baselines and methods aren’t published here. The work falls under confidentiality agreements with my employers. I’m glad to discuss the details under NDA.
Background
Software engineer who has worked up the stack from embedded firmware to mobile, cloud, machine learning, and now agent systems, at companies including Apple, Contran, AT&T, and Sam’s Club.