Amit Patole

Platform architect.
AI systems engineer.

My career started with a screwdriver, not a keyboard. In the late 1990s I was assembling and repairing PCs, everything from 486s to the first Pentium IVs, then reimaging desktops and running helpdesks for enterprise clients in Pune. It taught me the habit I have never lost: understand the machine from the metal up, and never call something fixed until you have watched it work.

From there I grew into servers and networks, and then into running a technical support department where I built Linux hosting platforms, automated site and database provisioning with scripting, and first learned that the real leverage in operations is removing the human from the repetitive path. At BMC Software I stepped up to enterprise platform administration, keeping high-availability and disaster-recovery environments healthy for the application suites businesses ran on.

Telecom is where I became a platform engineer. At Tech Mahindra I led the platform for a hosted carrier VoIP service on an OpenStack private cloud, working hands-on with the network functions that carry real calls for millions of subscribers. At VIAVI Solutions I became the primary cloud and platform architect for telecom systems deployed with operators across 23-plus production markets: Kubernetes and GitOps across multi-cloud and air-gapped environments, release automation that turned slow, error-prone deployments into something teams can rely on, and a security framework that audits every change. When production breaks, I am the escalation lead.

Working deep in AI platforms crystallized a question I had been circling for years: when a model produces an output, what does it actually mean for the work to be done? Not that a pipeline succeeded or a test passed, but done in the sense that you would stake your name on it. A 2026 paper I published explored one edge of that question, asking whether quantum optimization can route work across a multi-agent LLM cascade in ways classical greedy methods miss.

The open-source work below is the engineering answer to the same question. The Verel organism and its standalone companions are built on one conviction that has followed me the whole way, from the screwdriver to the model: nothing is done, and nothing compounds, until a grader returns a verdict.

Publications

2026

VANTAGE: AI Security Observability, Blast Radius, and Guarded Mitigation

Technical Design Document v1.0 · Zenodo, August 2026

10.5281/zenodo.21743977

2026

VR Time Travel: A Personal Nostalgia Engine with Real-Time AI Scene Correction

Zenodo, July 2026

10.5281/zenodo.21343168

2026

Attested Test Selection with a Verifiable Confidence Bound

Zenodo, July 2026

10.5281/zenodo.21231687

2026

Quantum-Enhanced LLM Cascade Routing: A QAOA Approach to Cost-Optimal Model Selection in Multi-Agent Systems

Preprints.org (MDPI AG), April 2026

10.20944/preprints202604.0413.v1 · Zenodo

2026

Grounded Cognitive Architecture for Enterprise AI Agents: Eliminating Hallucination via Mandatory Tool Execution at Scale

Zenodo, March 2026

10.5281/zenodo.19255196

Software

2026

latenzy: per-model LLM latency monitoring

Software release v0.2.0 · Zenodo, August 2026

10.5281/zenodo.21799202

2026

assaylab: validation intelligence for CI

Software release v0.3.0 · Zenodo, July 2026

10.5281/zenodo.21230859

github.com/amitpatole linkedin.com/in/amit-patole amit.patole@gmail.com

Verified AI, built in the open — a family of tools where nothing is “done” until a grader returns a verdict.

The Verel organism

Verel brain

Verified agents: nothing is “done” until a grader returns a verdict — verification, compounding memory, a fleet, agent-run CI/CD.

docs · GitHub

AgentVision eyes

Eyes for AI agents: a machine-graded visual feedback loop (DOM/contrast/OCR-grounded) agents consume to self-correct before claiming done.

docs · GitHub

Audel ears

Ears for AI agents: a machine-graded audio/voice feedback loop, grounded in a verdict on the produced sound.

docs · GitHub

Vitel vitals

Vitals for AI agents: graded observability returning pass/warn/fail on metrics, SLOs, and error budgets, grounded in a Metric.

docs · GitHub

Systems

Forge agentic delivery

A multi-agent system that builds what you ask for in your public repository and hands it over as a mergeable pull request — two unanimous policy judges, sandboxed tests it runs itself, a bounded repair loop, and a reviewer that is never the author.

case study · private repo

Observatory agent observability

Replay-grade tracing built for Forge: causality threads through graph state, prompt anatomy showing what truncation hid from the model, and forking of any recorded model call.

case study · private repo

Projects

latenzy llm latency

Per-model LLM latency monitoring — a Prometheus exporter and Grafana dashboards for Claude, OpenAI, and Gemini. Model A vs model B, not lab-level.

docs · GitHub

Tessera attested analytics

Attested NL→analytics for regulated finance: every number ships with the executed SQL, a runtime verdict, and a signed, auditor-verifiable receipt.

docs · GitHub

assaylab validation intel

Validation intelligence: attested test-selection with verifiable confidence bounds, failure-signature RCA, and CI analytics.

docs · GitHub

gapcert airgapped tls

Publicly-trusted TLS certificates for airgapped and egress-restricted fleets — an ACME DNS-01 broker, agents, and offline courier mode.

docs · GitHub

gapsign airgapped signing

Container-image signing for airgapped and egress-restricted fleets — sign at CI publish or mirror ingest, enforce at Kubernetes admission via Kyverno, all offline (no Fulcio/Rekor).

docs · GitHub