Mavera

An AI market-intelligence platform. I built it end to end, led the developers who joined, and turned its desktop agent into a product.

Role: Lead software engineer
When: Oct 2024 to Aug 2026
Where: Remote, Los Angeles
Stack: Next.js, Rust, MCP, Kubernetes

Mave answering a question about whether to start a SaaS business in 2026, with Answer, Evidence, Timeline and Flagged tabs above the answer. — Mave, the research agent, on iPad. Every answer comes with its evidence and a timeline.

The Timeline tab showing the triage, planning, research and execution phases of one answer, each with its duration in milliseconds. — The Timeline tab: the same answer, traced phase by phase with real timings.

The problem

Sheharyar Ishfaq was lead software engineer at Mavera from October 2024 to August 2026. Market research decides where media budgets go, but the classic tools are slow and narrow: pre-testing a video with a live panel could cost $50,000, and a customer journey map is often drawn from 15 or 20 interviews. The new shortcut, asking a chatbot about your audience, is fast, but it sounds just as confident when it is guessing.

Mavera models whole synthetic populations instead, so a team can test a message, a creative or a journey against the spread of a market rather than a handful of people. As its CEO, Jill Axline, puts it: “Decisions shouldn't move faster than evidence.” Every answer has to come with receipts: the sources behind it, the steps taken and how far to trust it.

What I built

The web platform

Next.js and Postgres, from first commit to version 1.14: accounts, workspaces and projects with eleven roles and per-project budgets, billing, and the backend for the iOS app.

Mave, the research agent

A triage step decides whether to ask a clarifying question; a planning step picks web search, news, SEO data and synthetic personas; then the answer is written with up to eight tool calls, every phase traced.

Pipelines

Eight Trigger.dev jobs that analyse videos and meetings, run synthetic focus groups, learn a brand's voice, keep the knowledge base fresh and write news digests.

The desktop agent

A Tauri app built on an open-source agent engine, which I turned into a product: Auth0 login, Stripe plans, credits, and a metered proxy in front of the model.

Infrastructure

Containers on Kubernetes (EKS) on EC2 behind a load balancer with a firewall, Postgres on RDS, S3, autoscaling and GitHub Actions deploys.

The team

Led the developers who joined as the platform grew, and wrote about half of the second codebase's commits myself.

How it fits together

Hover a layer.

Interface

The question is typed into the desktop agent, a native Tauri app I turned into a product, or into the web and iOS apps.

Rust
Tauri
React

Hard calls

Video jobs that pick up where they stopped

A long video is cut into chunks, each a 2×2 grid of frames plus its transcript. Finished chunks are kept, so a retry skips them instead of paying for them twice, and each chunk sees at most the five before it, which keeps the context inside the model's limits.

Retries that wait for the rate limit

Focus groups make many model calls at once. Instead of retrying blindly, the job reads the provider's rate-limit headers and schedules the retry for the moment the limit resets, and it stops at once when the quota itself has run out.

Every answer leaves a trace

The agent records each phase it goes through, with its duration and what it decided. That trace is the Timeline tab, part of the receipts Mavera promises: a researcher can see why Mave searched, which personas it used and where the time went before putting an answer in front of a client.

120+CompaniesIncluding Coca-Cola, Coinbase, Instructure, Morningstar, PitchBook, S&P Global and Spotify

$100K+Monthly revenueReached while I led engineering

2yLead roleFrom first commit to a team

His documentation is exemplary, both in precision and completeness, and it makes the entire build process more efficient.

Jill Axline, PhDCo-founder and CEO, Mavera