seekgb.com | LinkedIn | mcpindex.ai | ORCID
Product Manager blending strategy with execution. Developer platforms, public APIs, agentic AI, and MCP.
Developer platforms and the trust layer under agentic AI. The thesis in one line: a tool's contract can change after you approved it (no version bump, no republish event, nothing the consumer sees), so verification has to be continuous, not a one-time review at onboarding.
contract.pinned_at = approval_time // the snapshot, not the docs
served.contract != pinned.contract // drift, with no republish event
verification = continuous // advisory evidence, not a gate
I measure this at registry scale and publish the data openly. 3,500+ MCP servers observed daily; four datasets on Zenodo under CC-BY.
| Work | Status | Scope |
|---|---|---|
| OWASP GenAI Data Security Best Practices v2 | Authored Ch9 Pattern 8, "Tool-Contract Capture & Drift Verification"; amended the P7 Tier 1 contract-capture control | |
| OWASP FIASSE | Dependency stewardship for out-of-process dependencies. The framework's guidance covered code taken into a build and left out services called across a boundary; the project lead rewrote and consolidated the section around it (Discussion #20). Two commits merged to the framework document (PR #34). Runnable worked example: one REST dependency written twice, four tests, standard library Python | |
| MITRE ATLAS | Mitigation submitted for AML.T0104 (Publish Poisoned AI Agent Tool), covering post-approval contract mutation | |
| OWASP AISVS C9 | Real-data conformance fixtures for action-class and reversibility controls (reference repo) | |
| Adversarial ML for IoMT Security | Elsevier, Internet of Multimedia Things Security, third author (DOI) |
All published under Bharti, Gautam (ORCID 0009-0001-4448-1438), CC-BY-4.0. DOIs below are concept DOIs; they always resolve to the latest version.
| Record | Type | DOI |
|---|---|---|
| mcpindex Drift Report - Edition v1 (contract-drift corpus) | Dataset | 10.5281/zenodo.21449149 |
| mcpindex Source Liveness — Baseline v1 (reachability census) | Dataset | 10.5281/zenodo.21501867 |
| MCP Registry Drift Panel v1 (longitudinal observation panel) | Dataset | 10.5281/zenodo.21709945 |
| MCP Declared-Effect Coverage and Contract Binding v1 | Dataset | 10.5281/zenodo.21778281 |
| Registry Descriptions Go Stale Unevenly: An 89-Day Measurement | Preprint | arXiv:2608.00997, 10.5281/zenodo.21728369 |
AI Context, San Jose (September 23, 2026): a 20-minute session on measuring contract drift across the MCP registry.
Every tool here started as a proof-of-concept to solve friction I hit while shipping enterprise AI infrastructure. I prototype before I spec.
|
The agent-native index of MCP servers. Recommendation API + drop-in MCP server ( |
Enterprise-grade agent builder factory. A composable framework for building governed AI agents with structured tool access, memory, and orchestration patterns. Paste an OpenAPI spec, get a production-ready MCP server. Built to eliminate the boilerplate that slows agentic adoption, the same friction I kept hitting when onboarding teams to MCP at scale. |
|
Classifies user queries by complexity and routes them to the right LLM: fast model for simple questions, heavy model for complex reasoning. Cuts inference costs without sacrificing quality. Exposes latency, model, and cost telemetry per request. A privacy-preserving proxy between your application and your LLM. A fast model replaces PII with placeholders, the heavy model processes only clean text, and the response is unmasked before returning. Enterprise compliance without crippling AI capability. |
Extracts structured behavioral personas from writing samples: communication style, decision-making patterns, values, expertise markers, and 15 behavioral dimensions. Feed it emails, docs, or feedback; get a portable PERSONA.md that evolves over time. The cognitive fingerprint that demographics miss. Agentic workflows cost 10-50x more than the pricing page implies. Models what no other tool does: context accumulation across multi-step agent loops, tool call overhead, orchestration pattern multipliers, and prompt caching savings. |
All deployed and live at seekgb.com
Every tool here exists because I hit a real friction point. If nobody needs it, it is shelfware.
I build working proofs of concept to pressure-test an idea before committing engineering resources, because the best specs come out of building. That means discovery calls, support tickets, and using my own products. You cannot build for developers if you have never felt their friction firsthand.
I would rather validate with a rough working version this week than debate a polished spec for a month. Assumptions are cheap to hold and expensive to ship.
I've spent 12+ years building platforms at Microsoft (Power BI/Synapse, Windows Shell), T-Mobile, Trend Micro, and Qualtrics. The same thing keeps happening: powerful infrastructure exists and the people who need it cannot reach it. Each tool above is a piece I went looking for, did not find, and built.
Seattle, WA | GMT-07:00