AI-native consulting Β· Engineering lab
AI agents that run in production, designed and built end to end.
Zen Algorithms is an AI-native engineering practice. We design, build and deploy agent systems β multi-agent orchestration, MCP integrations, RAG, and the evaluations and controls that keep them dependable β for teams that need them to work in production, not in a demo. Our lab is where we build, run and evaluate the approaches we bring to client work.
Services
What we deliver
From one automated workflow to a multi-agent platform. Engagements range from an architecture review to production delivery, each with agreed outcomes and a proper hand-over.
Workflow automation
Prompt-chained pipelines for fixed, auditable processes β the right tool when the steps are known.
Multi-agent systems
An orchestrator coordinating specialist subagents for open-ended work, with human approval and escalation where it matters.
MCP integrations
Model Context Protocol servers and clients in Python and Java that connect agents to your databases, APIs and tools, with scoped permissions.
RAG and knowledge
Retrieval on PostgreSQL + pgvector: chunking by data shape, hybrid search (BM25 + vector) and re-ranking; Vertex AI Vector Search for managed scale.
Model routing and cost
One adapter layer across Claude, OpenAI, Gemini and Grok β portable to Bedrock, Vertex AI and Azure OpenAI β with routing, caching and fallbacks so cost tracks value.
Evaluation and guardrails
Golden-set tests and regression gates before any prompt or model change ships; PII scrubbing; policy enforced in code, not left to the prompt.
Ways to work with us
An outside look at your agent, RAG or MCP design: risks, cost, and a prioritised plan.
A locked spec and a working prototype on your data, so the decision rests on evidence.
End to end, from architecture to production, with evals, runbooks and hand-over.
Set your engineers up to work spec-first with AI coding agents (Claude Code, Codex, Gemini CLI), with the guardrails your security team expects.
How we work
- 1Scope
The business goal, the data, the constraints and what βdoneβ means β written down.
- 2Architecture
The architecture and acceptance criteria are agreed before any code. AI models critique the spec; Sam reviews their findings and makes the decisions.
- 3Build
Implementation with AI coding agents as everyday engineering partners, gated by golden-set evals.
- 4Run
Deployed on your cloud (Google Cloud, AWS or Azure), monitored, documented and handed over.
Selected work
Systems we have designed and built
The architecture write-ups are public, so you can judge the thinking before you talk to us.
Agentic insurance underwriting
Agents prefill each section in dependency order; underwriters approve per provision, and their decisions feed retrieval memory. Underwriting rules live in configuration rather than code.
Read the write-up βArchitectureMulti-user agent platform
Per-user memory, per-turn context assembly and a managed agent pool that scales against a policy.
Read the write-up βIn daily useMulti-model review
Claude synthesises while OpenAI, Gemini and Grok review in parallel β the same panel that reviews our own specs and code.
Read the write-up βLab Β· in productionAutomated content pipeline
The agent pipeline behind WatchAlgo: 1,600+ generated solutions, each validated by a 12-check report card before publishing; failures go back to the agent with the reason.
Read the write-up βThe lab
We build our own products the way we build yours
Live, used and open to inspection: the evidence behind our advice.
CosmicKeys
Live Β· typing platformTouch-typing lessons with animated hand guides and voice narration, on ISO and ANSI keyboards in 29+ layouts.
A multi-language, multi-region product with payments, running on Google Cloud.
WatchAlgo
Live Β· algorithm learning3,200+ problems with step-by-step visualizations and solutions in three flavours β technical, fun and spiritual β in JavaScript, Python and Java.
Content generated by an agent pipeline and checked by automated quality gates.
Mnemos
Open source Β· MITA local-first personal RAG: drop a folder, ask a question. Embeddings and chat run on your machine; everything lives in one SQLite file.
Private-by-default retrieval, released for other developers to run.
CosmicFlow Archives
Live Β· 15 archivesIndependent archives of full-length talks, podcasts and lessons β every session in order, in up to six languages, refreshed nightly.
One template, one persona file per archive, and search submission built into every deploy.
More architecture write-ups and experiments at sammuthu.com β
About
Who you work with
Zen Algorithms is the independent AI-native engineering practice and product lab of Sam Muthu, an AI architect with 15+ years of enterprise engineering: five years at Wells Fargo under SOX and PCI-DSS, including the move of an authentication platform serving tens of millions of users to microservices, plus Ericsson MediaKind, HP/Palm and Hartford Life. Hands-on with enterprise GenAI since 2023.
Designed for audits: SOX, PCI-DSS and telecom SLAs.
The person who designs your system also builds it.
Claude, OpenAI, Gemini or Grok β chosen per task, not per vendor.
Tell us what youβre building
A short note on the problem, the data and the timeline is enough to start the conversation.