Sarma Portfolio
A hire-me page backed by 25 public GitHub repos with real commit history, dependabot-verified security patches, and a live zero-knowledge AI chat product he built and shipped solo.
Location
Hemel Hempstead, Hertfordshire, United Kingdom
Skills
Notable
- 25 public GitHub repositories under sarmakska, with real multi-month commit histories, dependabot security patching, and open issues/PRs on the flagship repo
- sarmalink — a live, shipped BYOK AI chat product with client-side AES-256-GCM/PBKDF2 encryption and 17-provider failover, running in production at ai.sarmalinux.com
- SarmaLink-AI — an OpenAI-compatible multi-provider LLM gateway with 36-engine failover across 7 providers, MIT-licensed, 7 GitHub stars, 3 forks
What makes it work
A breakdown of the choices that make this portfolio stand out.
Every claim on the page links to checkable evidence
Each capability row ends in "Evidence:" followed by a specific GitHub repo, product page, or blog post — not a generic "see my work" link. Following github.com/sarmakska/Sarmalink-ai confirms the claims directly: 67 commits, a real commit log going back through 2026, dependabot-authored dependency bumps with green "Verified" badges, 5 open issues, 5 open PRs, a CONTRIBUTING.md and SECURITY.md. This is what an actively maintained open-source project looks like, not a portfolio repo created for show.
Self-scored capability matrix instead of a skills cloud
Instead of a wall of unranked tech-logo badges, the page scores eight capability areas out of 5 and says outright "not maxed across the board, because that would not be true." That single sentence does more credibility work than any amount of superlative copy — it reads as an engineer being honest about depth versus breadth rather than a marketer inflating a resume.
A live, working product as the flagship proof point
sarmalink at ai.sarmalinux.com is presented as the "densest piece of evidence... because you can open it in a browser right now" rather than a screenshot or a demo video. Client-side AES-256-GCM encryption with PBKDF2 at 600,000 iterations and multi-provider failover across seventeen LLM providers are specific, checkable technical claims, not vague "built an AI app" language.
Explicit, unglamorous constraints on what he wants
The page spends real space on "PAYE only, not contract, not Inside IR35" and "United Kingdom only" — details that filter out mismatched recruiters rather than casting the widest possible net. Combined with a personal blog that includes F1 write-ups and a Samsung phone review alongside the engineering essays, the whole site reads like an actual person's corner of the internet rather than an agency funnel dressed up as a portfolio.
What AI Agent Architects can take from this
Specific, actionable tips to apply to your own portfolio — no generic advice.
Link every capability claim to a specific piece of evidence — a repo, a PR, a whitepaper — instead of a generic "portfolio" or "GitHub" link at the top of the page. Specificity is what makes a claim checkable.
State a self-assessed limitation somewhere on the page ("not maxed across the board, because that would not be true"). One honest caveat does more for credibility than ten more superlatives.
If you have a live product, lead with the URL to the running thing, not a screenshot. A hiring manager who can click through in ten seconds trusts the rest of the page more.
Say plainly what kind of role and work arrangement you will and won't take (contract vs. permanent, location, seniority). It reads as confidence, not as narrowing your options.
Keep non-engineering personality on the site (hobbies, opinions, reviews) alongside the professional content — it is one of the more reliable signals that a site is a real person's and not a template.
Projects
A solo builder's stack for running LLMs in production: a live bring-your-own-key chat product, the open-source failover gateway underneath it, a durable agent orchestrator, and a memory layer for coding agents. Each repo is documented in full, as the hire-me page promises.
sarmalink
A bring-your-own-key AI chat workspace with automatic failover across providers and six modes: Smart, Fast, Reasoner, Live, Coder and Vision. Keys and chats are encrypted in the browser with AES-256-GCM, using keys derived through 600,000 PBKDF2 iterations. Coder mode builds complete apps in a sandboxed preview and saves them to the user's machine.
A shipped product with real client-side cryptography and multi-provider failover, built and run by one person.
Client-side AES-256-GCM encryption with 600,000-iteration PBKDF2
17 providers, free tiers first, automatic failover on rate limits
Installable PWA; Coder mode builds and exports full apps
A self-hosted, headless Next.js LLM gateway that routes every message through an ordered failover of engines. On a rate-limit or error it moves to the next engine in under 50 milliseconds. It speaks the OpenAI chat-completions contract, caches stable prompt prefixes across providers, passes tool calls through to MCP servers, and tracks cost per model.
It handles the unglamorous production parts of calling LLMs: failover timeouts wrapped in abort controllers, cross-provider caching, and cost accounting.
Ordered failover with a <50ms hop on 429/5xx
OpenAI-compatible contract plus a typed SSE event protocol
MCP tool-call passthrough and per-model cost tracking
A multi-agent workflow engine where workflows are typed graphs and state is saved to Postgres after every node, so a crashed run resumes from its last checkpoint. Agents get token, tool-call and wall-clock budgets that abort the run through an AbortSignal. A Next.js inspector draws the graph and steps through any run.
It treats durability, replay and hard budgets as first-class features, which is what agent orchestration needs to run in production.
Checkpoint after every node; deterministic replay from any step
Budgets that halt the run instead of logging a warning
Structural graph validation before any run starts
A local memory and observability layer for Claude Code that also runs as an MCP server in Cursor, Windsurf, VS Code and JetBrains. It turns each session into a durable summary, gives the next session a project knowledge feed at startup, and lets the agent read single symbols or line ranges instead of whole files.
It targets the two ways long agent sessions fail: running out of context and forgetting between sessions. Its efficiency claim comes with a benchmark anyone can re-run.
Cross-session memory distilled automatically from each session
Scoped reads by symbol or line range; ~95% fewer tokens per read (own benchmark)
Shared local bus so several agent tabs coordinate
More AI Agent Architect portfolios
Ready to build your portfolio?
Follow the AI Agent Architect roadmap — skills, projects, and timeline to get hired.