Skip to content
View SA-Ark's full-sized avatar

Highlights

  • Pro

Block or report SA-Ark

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
SA-Ark/README.md

Ark — Production AI Systems Engineer

typing

One engineer running a 120+ service production AI platform, mostly in Rust — multi-agent orchestration, RAG and search over vector databases, and GPU media pipelines.
I also design and build the front of house: fast, accessible, AI-native Next.js interfaces — an AI-enhanced design system on the front, llms.txt and structured data underneath so the sites are discoverable by LLMs and answer engines. The same person owns the model, the infrastructure, and the pixels.

Anyone can ship a demo. I build the system that's still standing eighteen months later.

ark.chakrakali.com  ·  engineering@chakrakali.com  ·  open to a few remote contract / consulting engagements


What I'd point you at first

  • portfolio — the map of the whole estate: the architecture, how one person runs 120+ services, and links to every live demo. Start here.
  • nexus — how I keep a fleet of AI agents from falling over. Work runs off a dependency graph, a stalled agent is caught by evidence rather than a timeout and retried, and anything genuinely risky stops to ask a human. Watch it run →
  • aegis — a Rust CLI that grades a codebase or a live URL the way I would on day one of a rescue: leaked secrets, shaky dependencies, untested critical paths, sloppy config. One binary, one score. cargo install aegis-audit
  • mindvault — the memory layer behind my agents. Hybrid recall (BM25 + vectors, fused) over a knowledge graph on pgvector, and no retrieval change ships without the eval harness proving it (recall@k, MRR, latency).
  • scour — the search primitives underneath that, pulled into a zero-dependency Rust crate: BM25, HNSW, reciprocal-rank fusion, and chunking that never splits a UTF-8 character. Live demo → · cargo add scour-search
  • crucible — an evaluation harness for LLM and RAG systems. Score what your model actually returned against ground truth — recall@k, MRR, nDCG, token-F1 — with pass/fail thresholds that drop straight into CI. cargo install crucible-eval

estate status: 120+ self-hosted production services, 12,000+ contributions in the last 12 months, 180K+ memories indexed at 100% embedding coverage, Lighthouse 100

metrics contribution calendar languages

contribution activity graph

3D isometric contribution calendar

How I work

The industry has gotten very good at selling the demo. A model that shines in a notebook, an agent that performs once on a launch stage — and then comes apart the first week it meets real customers, real load, and a real invoice. Dazzling in five minutes, unaccountable over eighteen months. That gap between the pitch and the operation is where most AI budgets quietly disappear.

My work lives on the other side of that gap. A system isn't finished when the code runs — it's finished when it survives real use on the live product, because standing something up is the cheap part and keeping it upright is the part a business is actually paying for. I don't claim a capability I can't show you running with the numbers behind it, so what you're buying is a measured outcome rather than a slide. And when something breaks — it always eventually does — I spend the time to find why instead of quieting the symptom, because the shortcut that saves an afternoon today is the customer-facing outage next quarter. What you're left holding is a system that grows more dependable the longer it runs, not one that decays the moment I walk away.

Selected design work

Interface design across industries — luxury retail, fintech, healthcare, SaaS, and more — each a distinct brand with its own identity, motion language, and layout system, designed and built end to end. Explore the live, interactive gallery →

E-commerce — AURUM Atelier luxury watch storefront
E-commerce · AURUM Atelier — luxury storefront + 3D configurator
Fintech — NORTHWIND private bank landing
Fintech · NORTHWIND — private bank, floating 3D metal card
SaaS — PULSEGRID product analytics dashboard
SaaS · PULSEGRID — product analytics with live charts
Healthcare — WILLOW Health clinic landing
Healthcare · WILLOW Health — calm clinic + booking flow
Dev Tools — TORQUE CI/CD pipeline landing
Dev Tools · TORQUE — CI/CD with a self-typing terminal
Music — VOLTA player landing
Music · VOLTA — lossless player with live visualizer
Travel — VELA Voyages editorial landing
Travel · VELA Voyages — editorial, film-backed itineraries
Real Estate — ATLAS & CO. property brokerage landing
Real Estate · ATLAS & CO. — curated listings + editorial home tours

Eight of twelve industries, each shown in motion · explore all twelve, interactive →


Stack

Languages · Rust · TypeScript · Python · SQL

AI / ML · LLM orchestration · multi-agent systems · RAG · vector search · knowledge graphs · pgvector · embeddings · GPU inference · MLOps

Backend / Data · Axum · Tokio · PostgreSQL · Redis · Qdrant · REST APIs · WebSockets · async / concurrent systems

Infra / Ops · Docker · Linux · systemd · Cloudflare · CI/CD · observability · self-hosted fleet · Model Context Protocol (MCP)

Frontend / Design · Next.js · React · Tailwind CSS · design systems · motion / micro-interactions · accessibility (WCAG) · responsive UI · AI-assisted design tooling · data visualization

AI-native web / GEO · llms.txt · JSON-LD structured data · generative-engine optimization · answer-engine discoverability · schema.org · semantic markup

tech stack

Pinned Loading

  1. portfolio portfolio Public

    120+ production services, one engineer. Architecture, case studies, and live demos of the estate.

    1

  2. aegis aegis Public

    Production readiness audit CLI — scan a codebase or probe a live endpoint for security, dependency, test-debt, and config risk.

    Rust

  3. mindvault mindvault Public

    Memory engine for AI systems — hybrid RAG + knowledge graph, architected for billion-node scale.

    1

  4. nexus nexus Public

    Multi-agent orchestration in production — live demo: watch 9 agents fan out, fail, recover, and ask a human.

    1

  5. scour scour Public

    Embeddable hybrid search primitives for Rust: BM25, HNSW, reciprocal-rank fusion, UTF-8-safe chunking. Zero dependencies.

    Rust 1

  6. crucible crucible Public

    Score LLM & RAG outputs against ground truth — recall@k, MRR, nDCG, F1 — with thresholds and a JSON report. Rust.

    Rust 1