Independent software engineer. Multi-provider AI gateways, agent orchestration, real-time voice, evaluation harnesses, infrastructure as code. Nineteen open-source repos. Eighty-three long-form posts. One client at a time, every commit on me.
By the numbers
live19
Open-source repos
MIT-licensed
83
Technical posts
with charts + citations
~5
Years in production
shipping software
< 50ms
AI failover handoff
across many engines
A live feed of the studio. Public releases, conference notes, code merges, the launches I am actually working toward.
Jensen opened with the largest single-keynote drop in years. Vera Rubin NVL72 in full production. RTX Spark and DGX Station GB300 making personal AI compute real on Windows. Cosmos 3 plus the Isaac GR00T humanoid reference. Nemotron 3 Ultra at 550B mixture-of-experts. Plus a fresh media stack and an enterprise agent runtime. Long-form open-source-engineer take with NVIDIA-courtesy images.
Read moreThree days in San Francisco, six themes that defined where AI engineering is going. MCP took the year. Multi-provider gateways became table stakes. Voice cleared the sub-second bar. Evals as code beat the dashboard era. Local-first assistants quietly rose. Full long-form write-up plus my own roadmap response.
Read moreecho is the open Jarvis. Brain-agnostic, voice loop, vision, persistent memory, translucent multi-monitor HUD, MCP skill bus. Runs on whichever AI subscription you already pay for. Cross-platform from one Rust core. MIT licensed. Signed installers for macOS, Windows and Linux. Watch the repo for the drop.
Read morePR #1 merged. The MCP stdio server now starts on Windows. Path separator normalisation in the entry guard. Zero behaviour change on macOS or Linux. Confirmed handshake returns all nine sp_ tools on a Windows build.
View on GitHubIntent auto-routing. Multi-step agent runner with SSE event stream. MCP-shape tool catalog. TTS and STT cascades with Cloudflare plus Gemini fallbacks. Live-data tools across weather, FX and HN with no API keys required. Image generation with rotation across CF account-token pairs. Quota tracker on a Supabase view. Smart-suggestions endpoint. Reasoning-leak stripper. Markdown to PDF plus JSON to XLSX export routes.
Read moreA long read on why MCP-shaped tools, multi-provider failover, journaled agents, evals-as-code in CI, sub-second voice loops, and local-first desktop assistants are now the assumed baseline. Honest opinions. Real citations. My own roadmap in response.
Read moreA disciplined replacement for the default coding-agent loop. Lossless context compaction, per-project memory that survives restarts, a live local dashboard for sessions, sub-agents and token spend. Built so the agent stops burning my tokens on things it should already know.
Replace Google Drive, iCloud, Plex Pass, and a paid password manager with one machine you already own. Add a local AI model on the same box for free. The full step by step guide, hardware specs, install, Immich, Jellyfin, Tailscale, Ollama, mobile setup, free vs paid.
Multi-provider AI backend with multi-engine failover. OpenAI-compatible proxy, persistent memory, image generation, live web search. Now with cross-repo plugin system and Manus integration.
Most agencies are running ten projects simultaneously. Your work gets squeezed between everyone else's deadlines. I do the opposite: I work with a single client until the job is done, then take on the next.
It means I'm expensive and selective. But it also means the work is consistently excellent, and you always know exactly where things stand.
Your project gets my full attention, not shared with four others
You talk directly to me, not an account manager
I know your codebase, your goals, and your constraints inside out
No briefing a junior developer who has never seen your product
Faster decisions, faster progress, better results
Six lanes I have shipped repeatedly enough to publish the underlying tools as open source. Each tile links to the matching service page and the open-source repo it draws from.
Multi-provider failover, OpenAI-compatible proxies, persistent memory, image generation, live tools. Single-provider risk in production is a pager waiting to happen.
Durable multi-agent workflows with deterministic replay, journaled state in Postgres, tool/token/wall-clock budgets, BullMQ queue, Inspector UI. Workflows that survive restarts and pass audit.
Sub-second WebRTC voice loops with pluggable STT/LLM/TTS adapters and barge-in. End-to-end RAG starters with cited streaming answers, ten-minute clone-to-ship.
Datasets as files, scorers as functions, traces in DuckDB, regression mode that fails CI when a release loses ground. OpenTelemetry across every workflow step.
Helm charts for Next.js with full observability stack preconfigured. Terraform composing Vercel + Supabase + Cloudflare + DigitalOcean. Eight minutes from empty cluster to fully observable.
Next.js App Router, TypeScript, Supabase Postgres with Row Level Security, Resend, Vercel. Server actions, ISR, OG images, JSON-LD, the boring tests passed.
I'm a solo engineer who partners directly with founders, CTOs, and technical leaders.
No account managers. No handoffs. Just focused, high-quality work.
When you work with SarmaLinux, you work with me, and I bring the same care to your project as if it were my own.
SarmaLinux is not only a technology studio. It is also a growing community of engineers, developers and builders collaborating on meaningful technology projects.
Connect with skilled professionals working on meaningful projects
Collaborate on AI systems and automation platforms
Build and test new software tools together
Learn from others and share your expertise
Members collaborate on AI systems, automation platforms, cloud architecture and experimental software tools. Anyone interested in building impactful technology is welcome to participate.
Join the CommunityI pick projects I would have loved when I was starting out. Tell me about yours and I will reply within a week. No payment, no contract, just real engineering help in my own time.
Every project I publish solves a problem I hit in real client work and decided to write down properly. All MIT licensed.
AI gateways, agent runners, voice loops, evaluation harnesses, RAG starters, OCR pipelines, MCP servers, a multi-tenant SaaS starter, a WebAssembly sandbox, a Raft KV store, an LSM-tree engine, and the infrastructure glue underneath them.
Long form, opinionated walkthroughs of the systems I run for myself and for clients. Real commands, real config, real warnings about what tends to go wrong.
How Sarmalink-AI routes a single OpenAI-shaped request across 14 backends, fails over in under 50ms, and never returns a 500. Health checks, intent routing, cost guards.
Read playbookMediasoup, Whisper, Groq, OpenTTS. The architecture I use for natural voice agents that respond in under a second on a £6 VPS.
Read playbookA real Model Context Protocol server: FastAPI, tools, resources, auth, deployment, observability. Built on the mcp-server-toolkit pattern.
Read playbookBringing an old datacentre GPU back to life for local LLM inference. Cooling, driver gotchas, model picks (Gemma 3 12B, Llama 3.3, Qwen 2.5 Coder), real tokens-per-second.
Read playbookWhen the Vercel bill starts to bite, here is the exact recipe to move a production Next.js 16 app to a self-hosted VPS without losing any features.
Read playbookReplace S3 in your stack with R2: zero egress fees, S3-compatible API, signed uploads from the browser. The pattern I use for invoices, blog images, contracts.
Read playbookPricing tiers, customer portal, webhooks, recurring usage. The smallest amount of code that gets you a real, live, SCA-compliant billing flow.
Read playbookThe exact magic-link flow I use on the admin and client portals: server actions, cookie session, callback handling, RLS hardening. With the four mistakes I made first.
Read playbookDaily pg_dump, point in time recovery, off site to R2, monthly restore drills. Most teams have backups. Far fewer have backups that work.
Read playbookWhat to put in Terraform, what to leave in the dashboard, and how to keep the diff small. The exact stack I use for Vercel, Supabase, Cloudflare, DigitalOcean.
Read playbookReplace Google Drive, iCloud, Plex, and a password manager with one box you already own. Hardware, install, Immich, Jellyfin, Tailscale, local Ollama AI, free vs paid.
Read playbookA production-ready n8n install on a £6/mo VPS, Docker, Caddy, Postgres, backups, the lot. The setup I run for myself.
Read playbookWiring Supabase to a Next.js app on Vercel without the auth foot-guns. RLS, server actions, type-safe queries.
Read playbookCutting a £400/mo Zapier bill to £6 without breaking the business. A real migration plan with patterns and pitfalls.
Read playbookA working OCR receipt to JSON pipeline using Claude vision, with the prompt and post-processing.
Read playbookThe pattern I use to ship schema changes in production without a maintenance window. Add, deploy, backfill, switch, drop.
Read playbookThe five SaaS subscriptions I cancelled last year, and what they were replaced with. Real numbers, real trade offs.
Read playbookTechnical writing on AI, systems, and engineering decisions.
A long, opinionated recap of the themes that defined the AI engineering field at the start of June 2026. MCP as the universal skill protocol. Multi-provider gateways as table stakes. Voice loops as a serious production stack. Evals as code becoming mainstream. The quiet rise of local-first assistants. And what I am shipping in response.
14 min read min readAI EngineeringJensen Huang opened GTC Taipei at Computex 2026 with the largest single-keynote product drop NVIDIA has done in years: Vera Rubin NVL72 in full production, the new Vera CPU, BlueField-4 STX, RTX Spark, DGX Station GB300, Cosmos 3, Nemotron 3 Ultra, Jetson Thor with Isaac GR00T, plus the Agent Toolkit and OpenShell runtime. This is the recap for an open-source AI engineer trying to decide where to put the next six months of effort.
13 min read min readSelf-hostingThere are five reasonable choices for a home NAS operating system. I run ZimaOS on my own box. Here is why it beat the others for me, and the kind of person it does not suit.
8 min read min readTechnology should not only power businesses, it should also help communities.
Helping charitable organisations operate more effectively with modern tools
Supporting education programmes with digital platforms and automation
Enabling community groups to focus on their mission, not administration
SarmaLinux supports initiatives where modern technology can help charities, education programmes and community organisations operate more effectively.
Automation systems, AI tools and digital platforms can help organisations focus on their mission instead of administrative work. Organisations working on meaningful causes are welcome to reach out.
charity@sarmalinux.comYou've got a problem. I solve problems with software for a living.
The fastest way to find out if we can work together is to talk.
Stack I build with