The architect behind the sovereign systems.
I'm Akshay — I design and build the AI systems Virohana stands for: agentic automation, autonomous self-healing infrastructure, and durable memory, all running on hardware the owner controls. This page is for people deciding whether to work with me.
I run Virohanalife OÜ as a one-person studio in Tallinn. The throughline of my work is making frontier-grade AI capability something you can own — systems that run locally, hold their own memory, recover themselves, and never quietly hand your data to someone else's cloud.
In practice that means an AI workforce wired into real software: agents that build, test, and ship under supervision; a memory layer that makes those agents improve across sessions instead of forgetting; and an autonomous engine that watches the whole thing, heals what breaks, and gates anything consequential behind a human. Most of the work is deterministic code and small local models — I reserve frontier reasoning for the parts that genuinely need it, which keeps it fast and cheap.
I care about intellectual honesty over optimism, verifying ground truth before trusting it, and doing a thing once and then making the machine do it forever. If that resonates, the fastest way to interrogate my work is to point your own AI at it — below.
Connect your AI, ask it anything.
My portfolio isn't a static page — it's a live, read-only server your own AI assistant can connect to and query. Ask about specific projects, my engineering judgment, how I'd fit a role, or have it pull my résumé. It's the same kind of sovereign, isolated system I build for clients, pointed at myself.
- In Claude (or any MCP-capable assistant), open Settings → Connectors → Add custom connector.
- Paste the server URL below and connect.
- When asked, paste the access key to authorize. Then just talk to your assistant — “tell me about Akshay's work,” “is he a fit for a staff AI engineer role?”, “show me his résumé.”
Two tiers of access: a standard key for recruiters and quick evaluation (below), and a technical key for engineering buyers who want the deep dive — issued on request. Read-only and isolated by design. Works with Claude, ChatGPT (Connectors), and IDE clients like Cursor / VS Code via npx mcp-remote https://hire.virohanalife.com/mcp.
https://hire.virohanalife.com/mcp
0d8ee5ce785b.hlvXU1Zm2ooBviD4HjdkBpqSXLnkMJmF
No account needed. Ask about projects, fit, and perspectives, or pull the résumé →
A separate, technical-scoped key unlocks the deep tier: detailed case studies of the hard problems I've solved, how I handle secrets and security, and architecture deep-dives — provisioning, observability, failure recovery, and more.
Issued per requester, and deliberately not published. The OAuth + scoped-key handshake you'd connect through is itself part of what's being evaluated — requesting one takes about thirty seconds.
Secrets you can't leak, keys you can't extract.
How I treat secrets and access on every system I build — and on the very server you can connect to above. Methodology, not topology: the specifics of any production system stay private by design.
Encrypted at rest (SOPS)
Secrets live as ciphertext, encrypted with SOPS (age / PGP). What's on disk or in a repo is never plaintext; decryption happens only in memory, at the moment of use.
Hardware-backed keys
High-value private keys are held in Apple's Secure Enclave and passkeys — generated in, and never leaving, the secure element. No extractable copy on disk to steal.
Scoped, short-lived tokens
Access is granted as narrow, expiring tokens, not standing credentials — the same two-tier scoped model behind the portfolio server above. Secrets never touch logs or build output.
Isolated by design
Powerful surfaces are never bridged to the open internet; public services run egress-isolated in hardened containers. The proof is upstream — that connector can't reach anything but a curated file.
Rotation, and anything touching secrets, access, or money, is a human-gated step — never autonomous. The machine proposes; the human disposes.
Capability, kept sovereign.
Agentic systems
Multi-agent workforces that build, test, review and ship under supervision — with cost-tiered routing so deterministic code and local models do the bulk, and frontier models only the irreducible reasoning.
Memory & retrieval
Durable, bi-temporal memory with semantic + graph retrieval, so agents carry context across sessions and improve instead of forgetting.
Autonomous infrastructure
Self-observing, self-healing systems: snapshot-and-revert, executable health checks, and auto-recovery — with consequential actions gated behind a human.
Sovereign deployment
Local-first, isolated, OAuth-secured services on hardware the owner controls. No metered clouds in the loop; the operator holds every key.