Experience

Where I’m building now, and the work that came before. Resume (PDF) ↗

Chert ↗

October 2026 to now

Infrastructure Engineer

Chert gives AI agents their own iMessage and FaceTime lines. I work on the Mac fleet that places and answers the calls, and the tools developers use to put an agent on a line.

Backed by Y Combinator (P26).

  • Replaced the physical FaceTime audio loopback with a software audio route in three days. It is flat from 100 hertz (Hz) to 10 kilohertz (kHz), where the hardware rig lost 18.7 decibels (dB) above 1.2 kHz, and about 15 milliseconds faster.
  • Built Chert MAU, the monitor and updater for the Mac fleet: Slack alerts, one-click healing of stuck lines, and signed releases that update a Mac without dropping a live call.
  • Shipped the chert command-line interface (CLI), where one command puts a LiveKit voice or video agent on a FaceTime line. 110 pull requests merged in my first four weeks.

Velofy ↗

April 2025 to now, Delhi

Co-Founder

An applied AI and security lab. We study how models find and fix vulnerabilities, build the harnesses and tools agents run on, and publish the work as open source projects and research.

  • glep, indexed code search for coding agents: 17 to 52 times faster than ripgrep on rare-pattern searches of the Linux kernel, with ripgrep-compatible output.
  • curl_reap, a Python scraping library with its own Transport Layer Security (TLS) 1.3 stack that sends Chrome’s handshake byte for byte, at up to 1.8 times the throughput of a C-accelerated rival.
  • Trove (repository index and memory for Claude Code), Querion (a read-only data analyst) and terbium (a document parser that calls a model only when unsure).

FableRoom ↗

February to November 2026

Software Development Engineer 2 (SDE 2), backend and applied AI

Sole backend engineer for FableRoom, a direct-to-consumer (D2C) home decor brand, and Trampoline, its business-to-business (B2B) arm, on one Django platform: 1,400+ commits across 15+ production systems.

Backed by Z47 (formerly Matrix Partners India), WaterBridge Ventures and Alteria Capital.

  • Built Maestro, the operator agent the company asks in Slack and on the web. It answers questions across Postgres, live Shopify, Mintsoft and Odoo, reads whole Slack channels, draws charts, and plans and creates backorders. Reads are open to everyone and every write waits on a one-time Approve card. It runs on a multi-model harness (Claude, Gemini, DeepSeek, GLM and Sarvam) with up to four parallel reads per turn.
  • Automated stock and orders across Mintsoft (the third-party logistics, or 3PL, warehouse), Shopify and Odoo: container stock sync, allocation, automatic backorders and carrier tracking.
  • Put every agent behind a read-only Structured Query Language (SQL) firewall, cut agent latency with 34% smaller prompts and parallel tool calls, and moved every credential to environment-only configuration across the codebase and its Amazon Web Services (AWS) deploys.

Fyle ↗

October 2022 to March 2025

Member of Technical Staff 2, integrations

Fyle is an expense platform for finance teams. Sage acquired it in 2025.

Backed by Tiger Global, Steadview Capital, Freshworks, BEENEXT and Pravega Ventures.

  • Built the integrations that post Fyle expenses into Sage Intacct, Sage 300 Construction and Real Estate (CRE), QuickBooks and Xero as journal entries against each company’s chart of accounts.
  • Owned both sides of each integration: the Django REST Framework backend that moved the data, and the mapping screens finance admins used to map and export it.
  • Sped up integration APIs with object-relational mapping (ORM) query tuning and PostgreSQL indexes, and moved approvals and notifications onto Celery background jobs.

First principles

Take away the job titles and most of my work comes down to one problem: data crossing a boundary between systems that do not agree. Three things follow from that, and they shape how I work.

The mapping is the product. At Fyle, an expense had to arrive in Sage Intacct, QuickBooks or Xero as a journal entry against that company’s chart of accounts. The two systems do not share a data model, so the backend could only move what the mapping said. That is why the work covered the mapping screens as well as the backend.

Fix the contract before the code. At Velofy the other side of the boundary is often a model, reading documents, web pages and databases it did not write. So I decide first what goes in, what comes out and what it may never do, then write the code that enforces it.

Make correctness checkable where data crosses. A confidence score on every extraction, a signature check on every inbound webhook, one method that guards every change to an order’s state. Each turns “it should work” into something a test or a person can verify.

Let’s connect

For hiring and opportunities, read my resume (PDF), book a conversation or write to hey@anishfyi.com.