Avinash Maurya
Available for consulting and forward deployed work

I ship platforms, integrations and AI features that hold up in production.

Engineering leader and hands-on builder with nine years in software. For the last five I've owned engineering for a B2B talent marketplace end to end, growing it from a single server to a multi-environment cloud platform while the talent network grew from 10,000 to 70,000 people.

See the work ↓How I can help
Avinash Maurya
Avinash MauryaEngineering leader · India · remote
10k → 70ktalent on the platform
~6,000commits in my current role
1 → 30+from one server to 30+ production workloads
9 yrsbuilding software, 4 products owned
01Platform scale · 2019–26Live

From one server to a cloud platform, while talent grew 7×

Starting point
A single EC2 instance running app, API and background work, deployed by hand, open to the internet. Fine for a few thousand profiles, fragile for anything more.
What I led
Four stages, each driven by a real limit we hit: serverless for the API, frontends on a CDN, then Kubernetes with GitOps and infrastructure as code across dev, UAT and production. Along the way we moved DNS and edge traffic to Cloudflare with a WAF in front, replaced the VPN with Cloudflare Zero Trust so internal tools and databases sit behind it, and added an ELK stack with APM for logs and tracing. No big-bang migration, no downtime for the product.
Result
The talent network grew from 10,000 to 70,000 people on a platform running 30+ production workloads on cost-efficient spot capacity, with a protected edge, private internal services, and logs, metrics and alerts the single server never had.
7×talent network growth, 10k to 70k, on the same small team
AWSServerlessKubernetesGitOpsTerraformCloudflareZero TrustELK
platform evolutionsimplified
  1. 2019–21One server App, API and jobs on one EC2 instance. Manual deploys.
  2. 2022–23Serverless 19 services behind one API gateway, queues and schedules. Frontends on a CDN.
  3. 2024–25Kubernetes + GitOps Managed Kubernetes on spot nodes, infra as code, three environments, backups, alerting.
  4. 2026AI-ready Identity service, AI connector, edge-hosted beta releases, website on the same cluster.
EdgeCloudflareDNS, proxy and WAF in front of every public host
AccessZero Trust VPNInternal tools, admin and data reachable only through Cloudflare Zero Trust
ObservabilityELK + APMCentral logs, request tracing, metrics and alerts to the team
Talent networkpeople on the platform
then
10,000
now
70,000
02AI integration · 2026Live

An AI assistant connected to a live marketplace

Problem
The account team jumped between several tools and spreadsheets to answer simple questions about talent and placements.
What I built
I sat with the team to learn the questions they really ask, then built a role-based MCP connector between AI assistants and the platform. Each user type gets its own tools, and access follows the same sign-in and permissions as the product.
Result
Talent search and account history in plain English, used to prepare client account reviews.
MCPPythonRole-based accessSearch
assistant · marketplace connectorillustrative data
Who do we have with private equity due diligence experience, available next month?
talent_search(skills: ["due diligence"], sector: "private equity", available: "next month")
✓ role: account team · scope allowed
14 matches. Top three by fit:
RK
Transaction services, ex-Big FourPE buy-side diligence · 12 yrs
from 1 Nov
SM
Interim CFOPE portfolio companies · 15 yrs
now
AD
Commercial due diligenceMid-market deals · 8 yrs
from 15 Nov
Ask about talent, clients or placements…↵
A recreated example of the kind of question the team asks. Names and profiles are made up.
03Architecture · 2026Live

Rebuilding the products on a modern stack

Problem
The client and talent experiences sat on an ageing frontend wired straight into a growing sprawl of backend endpoints.
What I built
I took ownership of the rebuild and became its main contributor: typed frontends, one API layer in front of the older services, and a dedicated identity service. Every hard call went into a decision record.
Result
A codebase new people, and AI agents, can find their way around, with the why written next to the what.
ReactTypeScriptNode.jsIdentityADRs
architecture · before and after
Before frontend
After client app talent app API layer auth
ADR-001When the API layer may read data directlyaccepted
ADR-002A dedicated identity serviceaccepted
ADR-005AI connector inside the platform, not beside itaccepted
Simplified. The decision records are paraphrased.
04Data quality · 2025–26Shipped

Skills data clean enough to trust

Problem
Fewer than half of the skills people added to their profiles could be approved automatically. The rest piled up in a manual queue, and search suffered for it.
What I built
Smarter autocomplete with alias matching, so "FM" and "financial modeling" land on the same skill. Bulk-added the skills people kept typing that we didn't have, and fixed multi-word search relevance before a relaunch.
61%auto-approval rate, past the 50% target
Search relevanceTaxonomyUX
profile · add a skillillustrative
financial mod
  • Financial Modellingexact
  • Financial modelingalias → Financial Modelling
  • FM (financial model)alias → Financial Modelling
  • Financial model reviewrelated
before
<50%
after
61%

Share of new skills approved automatically. Dashed line: 50% target.

05Platform · 2022–26Live

One path to production

Problem
Every service deployed differently, releases depended on who was on duty, and config drifted from what was actually live.
What I built
A serverless backend that grew to 19 services, then Kubernetes with GitOps for long-running workloads, all on infrastructure as code. One CI/CD pipeline with environment gates that match how QA and product sign off. Also moved edge traffic to Cloudflare and remote access to Cloudflare Zero Trust, with internal services kept off the public internet.
Result
Predictable releases, one place to look, and infrastructure that matches its code.
AWSKubernetesGitOpsTerraformCI/CD
release pipelineillustrative run

Build ✓

lint & types
unit tests
image

Develop ✓

deploy
smoke tests
auto-promote

UAT

deploy
QA sign-off
gate: QA

Prod

deploy
release notes
gate: product
serverless servicecontainer workloadsame pipeline for all of them
06Security · recentResolved

Supply-chain incident response

Problem
A compromised third-party account pushed malicious code into one of our source repositories.
What I did
Contained it, restored clean code, cut and tightened access, and wrote the leadership and team updates. The same playbook has since covered a brand-impersonation takedown and an email deliverability investigation.
0malicious changes reached production
Incident responseAccess controlCommunication
incident write-up · timeline
severity: highscope: one repositorystatus: closed
  1. DetectUnexpected commits spotted across many branches
    step 1
  2. ContainAccount access revoked, pushes frozen
    step 2
  3. RestoreClean history recovered and verified branch by branch
    step 3
  4. CommunicateLeadership and team briefed, partner notified
    step 4
  5. HardenAccess policy tightened, follow-ups tracked to done
    step 5
07Developer tooling · 2026In use

Agentic engineering with guardrails

Problem
AI-assisted coding is fast at first and expensive later. Unstructured prompting doesn't scale across a dozen repositories.
What I built
Agent instructions in every repository, a library of reusable skills, and a lockfile so shared skills are versioned like any other dependency. CI fails when a critical skill update is missed.
Result
A small team that ships like a bigger one, without losing track of what the agents are doing.
Claude CodeAgent skillsCI
repository layout + CIillustrative
repo/ ├── AGENTS.md # how to work in this repo ├── skills/ │ ├── release-notes/ │ ├── db-migration/ │ └── incident-writeup/ ├── skills-lock.json # pinned versions └── src/ $ skills audit --fail-on critical ✓ 3 skills pinned, all current ! 1 minor update available (release-notes) ✓ no critical updates · build continues

More work

SearchIn design

Semantic talent search

Hybrid keyword and vector search with a cost model. My own adversarial review caught a data-integrity risk before any code was written.

Design2026

Design system as code

Tokens, component specs and flows in one repository that designers, engineers and agents all read from.

Delivery2022–24

Metrics in engineering

NPS, story points and bug trends brought into planning, so roadmap calls are made with numbers.

Earlier2020–22

30+ services to serverless

At a previous company: moved 30+ services to an event-driven serverless setup and wrote CI/CD for 20+ of them.

Still hands-on. My commits per year in my current role, from the git history. The dip in 2024–25 is when I spent more time leading than typing. 2026 runs to mid-October.

05001,0001,500 5591,4611,756 5452011,436 202120222023 202420252026*
Services

How I can help

Primary

Forward deployed engineering

I embed with your team or your customer's, learn the real workflow, and build the integration, pipeline or AI feature that makes your product work there. Then I hand it over with docs and runbooks.

AI

AI integration and agents

Connect your product to AI assistants with MCP and proper access control, and set up agentic development that stays under control.

Review

Platform and cloud review

A focused look at your cloud, edge, access, observability and CI/CD. You get a written report with priorities.

Leadership

Fractional engineering lead

Architecture calls, delivery process, hiring help and partner oversight for teams without a CTO.

  1. Short callThe problem, who it's for, what done looks like.
  2. One-page scopeGoals, approach, timeline, risks.
  3. Build alongside youWeekly demos and a running decision log.
  4. Hand overCode, docs and runbooks your team can own.
About

Builder first, lead second

Avinash explaining a system diagram on a whiteboard

I started as a freelance frontend engineer at a talent marketplace in 2021 and grew into owning its engineering: frontend, backend, cloud, security and the team. Over that time we went from one server to a multi-environment cloud platform, and from 10,000 to 70,000 people in the talent network. I still write a lot of the code, and I also wear a product hat, so the why stays connected to the how.

I write things down. Decision records, runbooks and bottom-line-up-front updates are how I keep a small team fast and a client informed.

Previously at RAx Labs, SalesHandy and Burning Glass Technologies. M.Tech in Computer Science from NIT Arunachal Pradesh. Full history on LinkedIn.

Contact

Have something that needs to work in the real world?

Send a short note about the problem: what you're building, who it's for, and where it's stuck. I reply within two working days.

LinkedIn ↗