Sanchay Pahalwani

UTC+5:30

Sanchay Pahalwani

Engineering Manager at Mira Network who still ships every day

LLM agents, RAG and evals in production, shipped end to end.

Bengaluru, India (UTC+5:30).

Email GitHub LinkedIn

[ TELEMETRY ]

SOCKETS.CONCURRENT100,000+

simultaneous socket connections kept healthy across load-balanced servers (BrowserStack)

UPTIME99.99%

uptime after shipping the fault-detection service (BrowserStack)

INFERENCES.DAILY10M+

inferences validated per day by the AI evaluation core I helped build early components of (Mira)

REVENUE.Q1$1M+

first-quarter revenue from one geolocation compliance feature I built (BrowserStack)

TEAM.ENGINEERS7

engineers on the team I run (Mira)

YEARS.SHIPPING8+

years shipping production systems

[ SELECTED WORK ]

how-i-use-ai

A 4,110-line agent config failed and was replaced with layers that load only when relevant: rules with explicit triggers, skills for reusable debugging knowledge, a permissions allowlist, per-project config, and CI for anything that must hold no matter what the model decides. Every feature starts with a dated spec and plan in the repo, and agent work comes back as a report. Drawn from a team project under NDA: 331 commits in 7 weeks, 5 contributors, 16 plan documents and 9 design specs.

github.com/SanchayPahalwani/how-i-use-ai

deploy-to-babou

Agentic CI/CD in pure TypeScript/Node: every merge to main hands the release to a headless Claude agent that drives an external MCP server to draft a launch campaign, with a never-publish safety invariant enforced three ways and zero-credential offline verification.

github.com/SanchayPahalwani/deploy-to-babou

BlockWorks

Visual data-pipeline SDK and runtime: a single-call defineBlock() TypeScript API with end-to-end type inference, a columnar Rust engine bridged via napi-rs, and an AI block generator that uses the SDK types as the entire LLM context.

github.com/SanchayPahalwani/BlockWorks
BlockWorks canvas running a 50 million row Polars pipeline with live per-block row counts and an output table
FIG. 01 / PIPELINE CANVAS, 50M ROWS

veritas-eudr

Geospatial compliance backend: typed Python/FastAPI over PostGIS/PL/pgSQL with a measured ~103x GiST-vs-sequential query speedup and reproducible benchmarks.

github.com/SanchayPahalwani/veritas-eudr
Plot risk map marking deforestation risk per plot with an open popup showing one plot's validation disposition and evidence records
FIG. 02 / PLOT RISK MAP, EVIDENCE TRAIL

colonist-resource-tracker

A card tracker for colonist.io, the game I actually play.

Chrome MV3 extension that reconstructs the full public economy of a colonist.io game (bank, every player's hand, dev card counts) by folding the rendered game log through a MutationObserver and an ordered regex pattern table. Hidden information stays hidden: robber steals are tracked as explicit unknowns. Survives the site's SPA navigation with a watchdog that rebinds when the log remounts.

github.com/SanchayPahalwani/colonist-resource-tracker
Resource tracker overlay showing live bank and player hand counts during a four player colonist.io game
FIG. 03 / RESOURCE TRACKER, LIVE GAME

Earlier open source

CRUX Protocol (github.com/cruxprotocol): open-source identity layer for crypto wallets; designed the plugin architecture, wrote the reference implementation and developer docs, and supported the integrator community; adopted by 2 wallets, Edge and Cake Wallet.

[ HOW I RUN THE TEAM ]

EVALS

The release gate. Prompt, model and tool changes ship only after a 600-case regression set built from production conversations passes. Each week's failures and handoffs feed back into it; model-graded rubrics run at 0.81 agreement with human labels.

ESCALATIONS

Teach the agent. A human's answer to an escalation becomes a correction the agent reuses, so a question escalates once. Autonomous resolution went from 50% to 65% and repeat escalations from 25% to 5%, with no prompt or doc edits.

AI CODING

Agents across the team with shared rules and review norms, and a monthly session on what they're good at and where they fail. Extended from my team to all of Mira engineering, with MCP servers for internal tools.

HIRING

An interview loop with role scorecards, a system-design round and a debrief rubric; a hiring plan and levelling for SDE1, SDE2 and Senior. 6 hires, 2 promotions, 100% offer acceptance.

RELIABILITY

On-call with runbooks and blameless postmortems; error budgets per service and a weekly reliability review for all of Mira engineering. Deploys went from 2 a week to 6 a day; MTTR is under 30 minutes.

PEOPLE

30 non-engineers in sales and support trained on the agents and AI tools. 1 pod with a lead who runs its planning, coached in 1:1s.

[ EXPERIENCE ]

Dec 2024 to present

Bengaluru

Engineering Manager, Mira Network

Run the B2B AI division: grew the team from 3 to 7 with 6 hires and 2 promotions. We ship a plug-and-play agent SaaS that works inside 40+ platforms (Python/FastAPI, Node/TypeScript, React/Next.js): Agno agents with a 13+ tool registry, Milvus RAG, and a 600-case eval set that gates every prompt, model and tool change. Each human escalation becomes a correction the agent reuses; autonomous resolution went from 50% to 65% and repeat escalations from 25% to 5%. Still hands-on: architecture, code review, code on deadlines and prod bugs.

Jul 2024 to Dec 2024

Bengaluru

Team Lead, Mira Network

Led 3 engineers building the first eval pipeline and early components of the eval core that verifies 10M+ inferences a day. Shipped an MCP agent on GoPlus APIs for real-time security intelligence within 2 months of MCP's release.

Sep 2021 to Mar 2024

Remote

Senior Software Engineer, BrowserStack

Local Testing connectivity: tooling and diagnostics for 100,000+ simultaneous socket connections, fault-detection service that pushed uptime to 99.99%, geolocation data-region feature worth $1M+ in its first quarter, 17% session-execution latency win.

Mar 2019 to Aug 2021

Bengaluru

Software Engineer, CoinSwitch

Authorization and security module for crypto and fiat withdrawals, KYC, and PIN resets; Docker + AWS ECS migration: 60% faster deployments, 20% lower server costs.

Jul 2018 to Mar 2019

Bengaluru

Software Engineer, Muse Wearables

Backend for the game Planet Watch (serverless cloud functions); later built an activity classifier separating five states from raw smartwatch sensor data.

[ EDUCATION ]

B.E. in Information Technology, Shri Vaishnav Institute of Technology and Science, Indore. 2014 to 2018.