Guru Charan Gupta V B
Founder & AI Engineer at Proof of Tech
Senior full-stack / AI engineer. 8+ years building production systems end to end: web platforms, LLM and media-model products, agent infrastructure on the edge.
Eight years building production systems, end to end.
I've spent the last eight years shipping production systems: user-facing apps, developer tooling, content infrastructure, and the CI/CD and edge infrastructure underneath them.
From 2018 to 2025 I worked at Komodo, an open-source blockchain platform whose core technology was acquired by Gleec in 2025. I held five roles there (documentation, support, QA/DevOps, frontend, and finally senior platform engineering), and that order is why I build the way I do: I watched systems fail from the docs, from the support desk, from CI, and from production.
Today I build AI products. I'm the lead frontend engineer and a core backend engineer at an early-stage venture building a real-time, multimodal generative-AI web application (streaming conversational AI plus image, video, and voice generation), and I run Proof of Tech, the consultancy this site belongs to, founded in January 2026.
One build under Proof of Tech is a hands-free voice agent: one click opens a continuous browser call, speech recognition decides when a turn ends so nobody presses a button, and talking over the reply aborts the model and speech calls mid-turn, dropping audio already in flight so it stops instead of speaking over you.
I've also built an agent and capability cloud on Anchorage, my own open-source agent-safety libraries. It gives a hosted agent a priced, metered, and governed way to call outside services: a control plane owns definitions and deployments, and a budget authorizer counts in-flight reservations against caps behind a kill switch. Its capability gateway exposes a REST API and a remote MCP server.
We believe that evals are the contract every deploy has to pass.
Every AI system I've watched die in production died the same way: the demo was the spec, and nobody had written down what "working" meant. Proof of Tech is built on the opposite premise. Define what the system must do before building it, gate every deploy on that definition, and hand the whole thing over when it holds.
The firm is deliberately small and runs one engagement at a time, so principal engineering is never the bottleneck. How an engagement actually runs is on the about page; what you own when it ends is at handoff.
This site is the working proof, and so is the tooling. I built the platform solo, and the approval, audit, and policy layer I build agents on is open source, with the code and a live demo below. The colophon documents the engineering.
What shipped, with dates.
- ◇ 2026
Governance for agent workflows
I open-sourced Anchorage, an enterprise safety layer built on Mastra, so an agent's outward actions wait for human approval and leave an audit trail. It carries role-checked approval queues, a record of every action, egress and write-permission policy, and durable execution on Cloudflare: a run suspends at an approval step, survives a server restart, and resumes from its snapshot. CI runs that on the real Workers runtime, and checks that a run cannot be resumed with a forged grant and that nobody can approve their own request. It ships as three npm packages, each published with a provenance attestation.
- ◇ 2026
Browser actions and retrieval, in the open
I wrote understudy and vectorless-rag solo, both under MIT. understudy is a model-free browser-execution substrate for AI agents: it drives a user's already-logged-in Chromium tab over the Chrome DevTools Protocol behind approval gates, with credentials resolved service-side so they never reach a model's context. vectorless-rag answers grounded questions over a local PDF corpus without embeddings or a vector database, routing each request through a LangGraph agent and expanding a document tree, so every citation resolves to a page instead of an opaque chunk.
- ◇ 2026
Shadow-mode compliance triage
I built my own alert-triage platform for anti-money-laundering workflows on Anchorage. It reads privacy-masked blockchain-monitoring and sanctions alerts and returns evidence-cited recommendations while taking no regulated action. A deterministic rules engine runs the mandatory hard stops before any model call, so a model can never recommend closure over a triggered stop, and a human analyst makes every binding decision.
- ◇ 2026
A payment gateway on a public testnet
I designed, built, and operated a complete stablecoin payment gateway end to end on a live public testnet, for a regulated fiat-backed stablecoin on an EVM ZK-rollup L2. On the merchant side I built hosted checkout, an SDK, a WooCommerce plugin, and signed webhooks. Underneath, an event-sourced invoice machine wrote to a hash-chained double-entry ledger, custody held payouts behind maker-checker approval, and KYT screening ran fail-closed at checkout and on every payout destination. I wrote the Solidity contracts and tested them in Foundry behind a gas-drift gate in CI. Under my own load test, a single instance held 2,500 concurrent checkout sessions.
- ◇ 2026
This platform
I built proofoftech.org solo: a hybrid retrieval system (BM25 + dense embeddings, Reciprocal Rank Fusion) streaming cited answers over the field notes, a human-in-the-loop content pipeline whose citation gate resolves every arXiv identifier against its live page before a draft may cite it, and a Lighthouse performance budget enforced in CI. That pipeline sits behind 115 published notes drawing on 374 arXiv papers.
- ◇ 2019
Agama incident response
When the 2019 Agama wallet supply-chain attack hit, I wrote the security advisory and the first user communications, then ran the fund-recovery workflow after the white-hat sweep of roughly $13M in assets. It restored holdings to thousands of affected users.
- ◇ 2018–2025
A public web estate at scale
I shipped and ran 10+ public web properties (120k+ monthly visits, 1M+ monthly Google Search impressions) and owned the frontends, the CI/CD pipelines, and the web-server configuration-as-code under them.
- ◇ 2019–2025
Consensus-critical infrastructure
I led a containerized notary-node stack that helped protect 15+ blockchain networks against 51% attacks: coordinated daemons across four regions, 52k+ notarizations a season, kept running through successive protocol updates.
Bring us a hard problem. We'll show you what we'd build.
The first call is a free 30 minutes. You'll leave with the first cut of the build/no-build path, the riskiest assumption, and the eval we'd use to test it.