MuhammadZain
Skills
Architecture & Security Stack
Featured Ops
Production Deployments
A stateless multi-model deliberation engine that routes queries through five LLMs and orchestrates adversarial cross-examination between them. Outputs a structured Contested Zone that pinpoints exact disagreement vectors instead of averaging them away.
Role
Sole architect and engineer — designed the multi-agent debate protocol, FastAPI orchestration layer, and stateless API-key-on-client security model from scratch.
Problem
Single-model AI outputs hallucinate with full confidence. Cross-checking manually across five providers is slow and produces no structured diff of where models actually conflict.
Outcome
Deployed to production. Users can paste any query and receive a structured adversarial report with exact conflict zones in under 15 seconds, with zero server-side key storage.
System Stack
A full-stack voting platform built around Ring Signature cryptography — votes are anonymous by math, not by policy. OTP-gated auth enforces one-identity-one-vote without storing any linkage between an identity and its cast vote.
Role
Full-stack — implemented the Ring Signature scheme, designed the tamper-evident audit ledger in PostgreSQL, and built the React voting interface with OTP auth.
Problem
Digital voting systems either sacrifice anonymity for auditability or auditability for anonymity. Most implementations rely on policy controls rather than cryptographic guarantees.
Outcome
Voter anonymity is mathematically provable, not just promised. The audit ledger is publicly verifiable without de-anonymizing any vote record.
System Stack
A LangGraph-orchestrated pipeline that fans out code submissions to parallel specialist agents covering security, architecture, and style — each powered by Gemini 2.5. Execution gates on human approval before merging agent outputs into a final structured report.
Role
Full-stack — built the LangGraph DAG, FastAPI streaming endpoints, HITL gate UI, and the Recharts telemetry dashboard tracking per-node latency and token spend.
Problem
Code review is sequential and bottlenecked by a single reviewer's domain depth. Security issues slip through architecture reviews and vice versa.
Outcome
Three specialist agents run in parallel, cutting single-domain review time by ~60%. The HITL gate catches pipeline regressions before they merge to the final report.
System Stack
Connection Hub
Interact or Book a Meeting
Arrange a Meeting
Want to discuss a project, architecture, or simply connect over a coffee? Let's find a time to chat.