Hi, I'm Ankit Sharma, a full-stack software engineer and freelance web & mobile app developer in Kathmandu.

I build the parts of a product where the money and the data have to be right.

At MeroGo I build event ticketing, payments and anti-abuse features across a TypeScript backend, React Native apps and a Next.js admin. Before that I built Kubesense AI's Data Explorer and Dashboards. I also maintain promql-validator, an open-source PromQL editor and validator on npm.

Ankit Sharma in Kathmandu
Kathmandu,
Nepal
Ankit Sharma, full-stack software engineer based in Kathmandu, Nepal
Ankit Sharma
simulated buyers through ticketing, zero failed requests
100k
tests passing in promql-validator
1,085
open-source packages on npm
5
client teams turned into one query experience
5

About

How I work

the short version

I like the features where the requirements are still moving and a mistake costs real money.

I'm based in Kathmandu and have spent 2+ years working remotely with product teams in the US and India: payments, inventory, abuse prevention, and the query tools engineers use to debug production.

At MeroGo I led event ticketing from the PRD onward and built the controls that stop rides being cancelled and settled off-platform. At Kubesense AI I turned requests from five client teams into one query experience. At Qontentt AI I led the MVP frontend as an intern and joined full-time through a pre-placement offer.

I studied Computer Science at Amrita Vishwa Vidyapeetham on a full-tuition Chancellor Scholarship.

  • 1

    Features, end to end

    From PRD and API design to React, React Native and Next.js screens, tests and release.

  • 2

    Correctness under load

    Transactional inventory, idempotent APIs, payment state machines and reconciliation, proven with real-database race tests.

  • 3

    Platform integrity

    Rules-based abuse detection with configurable enforcement and admin review, kept explainable on purpose.

  • 4

    Developer tools and AI

    Parsers and editor integrations, and retrieval-grounded LLM pipelines with citations and confidence scores.

Experience

The climb so far

Product engineering across mobility, observability and AI content apps, for teams in the US and India.

  1. MeroGo

    Dec 2025 – Present

    Full-Stack Software Engineer · Remote, USA

    MeroGo is a US super app built around ride-hailing. I work across the Node.js and TypeScript backend, the React Native passenger and driver apps, and the Next.js admin and organizer portals.

    • Led event ticketing from the approved PRD to a production release pipeline: a standalone TypeScript service with PostgreSQL as the single authority for seat holds and stock, adding ticket sales as a revenue line alongside rides.
    • Hardened the payment path with scoped idempotency keys, one Stripe PaymentIntent per reservation, refunds run as a saga, an append-only ledger and nightly reconciliation against Stripe. Canary deploys roll back automatically on payment alerts.
    • Investigated repeat cancellations used to take fares off-platform, then built driver and passenger prevention: pair risk scoring, weekly limits, wallet penalties and an admin review queue, all tunable from admin settings without a release.
    • Built the visual banner builder for in-app promotions (layout, typography, image styling, live preview), so the business team publishes targeted, scheduled campaigns without an app release.

    TypeScript · React · React Native · Next.js · Node.js · PostgreSQL · MongoDB · Redis · Stripe · Google Cloud Run · Terraform

    Case study
  2. Kubesense AI

    Jun 2025 – Nov 2025

    Software Development Engineer I · Chennai, India

    Kubesense AI builds an observability platform for metrics, logs and traces. I built the querying and dashboard experience on the frontend.

    • Built the Data Explorer and Dashboards features from scratch in React and TypeScript: visual and PromQL query modes, multi-query formulas, drag-and-drop panel grids, dashboard variables and JSON import/export.
    • Turned conflicting requirements from five client teams into a hybrid visual Builder and PromQL Code experience that met their needs and helped the company secure a proof of concept.
    • Cut chart load time by 20% by optimizing ECharts rendering of large VictoriaMetrics and ClickHouse result sets, and shipped core features 50% ahead of schedule.

    React · React Native · Next.js · TypeScript · ECharts · PromQL · VictoriaMetrics · ClickHouse

    Case study
  3. Qontentt AI

    Apr 2024 – May 2025

    Software Development Engineer (Intern, then Full-Time) · Remote, USA

    Qontentt AI is an AI-assisted content app for social media creators. I joined as an intern to build the MVP and stayed on full-time.

    • Led frontend development of the MVP as an intern, then joined full-time through a pre-placement offer (PPO) and managed Play Store releases.
    • Built an OpenAI-powered content assistant (captions, hashtags, in-app guide) and cut its API latency by 80% with a Redis caching layer.
    • Built GraphQL APIs for users' social data and integrated Instagram (Meta) APIs, OTP login and RevenueCat subscription billing.

    React · React Native · Next.js · Node.js · GraphQL · OpenAI API · Redis · RevenueCat

  4. Optiblack

    Oct 2024 – Jan 2025

    Software Engineer Intern · Remote, USA

    A remote internship working on a client product's onboarding, frontend quality, analytics tracking and deployments.

    • Rebuilt onboarding with two-step verification and better security, and optimized its APIs for 200% faster responses.
    • Revamped the frontend UI, resolved bugs 200% faster and delivered internal Mixpanel tracking 300% ahead of the timeline.
    • Resolved AWS EC2 issues, helped with Play Store deployments, and supported client communication and feature planning.

    React · React Native · Next.js · Node.js · AWS EC2 · Mixpanel · REST APIs

  5. DeepThought EduTech Ventures

    Sep 2022 – Dec 2022

    Software Developer and Product Intern · Hyderabad, India

    A product and engineering internship across the company's web and mobile apps.

    • Resolved over 50% of open frontend issues across web and mobile, improving stability, and contributed to a microservices-based task management system.
    • Turned 10+ Figma designs into React and React Native screens for a dynamic feeds page, and took part in product ideation.

    React · React Native · Node.js · Figma

Selected work

Things I've built

the ones I'm proudest of

The problem, the engineering decisions and what each one delivered. Each sketch on the right is a small working model of the system.

01MeroGo · Professional work

Event ticketing and ride integrity

A ticketing service for a ride-hailing super app, built so it cannot oversell, double-charge or lose track of money under load, plus the controls that stop drivers and passengers cancelling rides to settle fares off-platform.

  • PostgreSQL is the single authority for holds and stock, enforced by a CHECK constraint in the same transaction as every order change.
  • Over 1,000 automated tests, including concurrent-request races and failure injection against real Postgres, Redis and a Stripe fake.
  • On a production-shaped load-test stack, a simulated rush of 100,000 buyers sold all 10,000 tickets through the waiting room with no failed requests.

TypeScript · Node.js · Express · PostgreSQL · MongoDB · Redis · Stripe · Google Cloud Run · Terraform · Jest · Testcontainers · fast-check · k6 · React Native · Next.js

merogo.us

Problem

MeroGo wanted to sell event tickets inside its app, starting with a US concert with a fixed on-sale date. Ticketing handles money, has hard stock limits and gets traffic spikes at on-sale, so the PRD set three rules that must never break: never oversell, never lose money state, never double-charge or double-issue. Separately, some drivers and passengers were cancelling matched rides and finishing them off-platform to avoid commission.

My role

The PRD came from the engineering lead. I led the build: the ticketing service and its test suite, the admin, organizer and door-staff screens in the Next.js portal, and the ticket flow in the React Native app. I also owned cancellation prevention across the backend, both mobile apps and the admin portal.

Inventory and payments

  • Holds and sales change stock with conditional UPDATEs in the same transaction as the reservation or order row, behind CHECK (sold + held ≤ total). Orders that span several ticket types lock rows in a fixed order to avoid deadlocks.
  • Redis holds only a short-lived availability hint that lets a sold-out type reject bursts cheaply. It never admits a hold, so a Redis flush or outage cannot cause an oversell.
  • Every mutation needs an Idempotency-Key, scoped to caller, method and route and fingerprinted against the request. Only successful and deterministic client-error responses are replayed; money routes keep their keys in Postgres.
  • One payable order and one Stripe PaymentIntent per reservation. Payment state only moves forward, so a decline followed by a success still issues tickets, and a paid order that cannot be fulfilled becomes a durable automatic refund.
  • Refunds run as a saga: record the intent, call Stripe, then void tickets and return stock. A refund counts as done only when Stripe confirms it, and an append-only ledger records every money movement.

Architecture and key decisions

  • Money and inventory live in PostgreSQL; event content stays in the main backend's MongoDB.
  • One image runs as two Cloud Run services: a public API and a private worker. The worker accepts only Google-signed tokens from allow-listed service accounts.
  • Background work uses a Postgres job queue: jobs are enqueued in the same transaction as the state change and claimed with leases (FOR UPDATE SKIP LOCKED), with retries, backoff and dead-lettering.
  • Ticketing verifies the main app's access tokens but can never mint one: ES256 only, public keys only, with issuer and audience checks.
  • A waiting room and load shedding admit buyers at the rate the purchase path can sustain during a hot on-sale.

Reliability and testing

  • Integration tests run against real Postgres and Redis (Testcontainers) and fire concurrent HTTP requests in the same tick. After every test, invariant checks confirm there is no oversell, no paid order without tickets and no unrecorded Stripe refund.
  • Property-based tests (fast-check) cover money arithmetic, refund allocation and the order state machines. A contract test validates every Stripe request against Stripe's own OpenAPI mock.
  • Nightly reconciliation runs SQL and Stripe consistency checks, recovers the safe cases automatically and escalates the rest.
  • A 23-scenario chaos suite injects database, Redis and Stripe faults during live traffic. No run produced an oversell or a duplicate charge.
  • Canary deploys send 10% of traffic to the new revision and roll back automatically if money-state signals or error rates get worse. In a local drill with a deliberately faulty release, rollback happened about 5 seconds after the traffic split.

Cancellation abuse

  • Rules-based and explainable on purpose: an admin can recompute any score by hand. Pair risk is capped at medium, so a regular commuter who keeps matching the same nearby driver is not punished.
  • Weekly cancellation limits, wallet penalties, account suspension with a dispute path, and a pre-cancel confirmation that shows the penalty and asks for a reason and optional photo proof.
  • Chat messages are checked for off-platform signals such as shared phone numbers, named payment apps or 'cancel and pay me directly'. Matches feed an admin review queue and never block a message.
  • Thresholds and penalties live in admin settings, so operations can tune enforcement without shipping a release.

Results

  • Load tests on a production-shaped stack (test traffic, not production): 40 tickets per second sustained on a single ticket type. With the waiting room on, a 100,000-buyer rush sold all 10,000 tickets in about 7 minutes with no failed requests and no oversell.
  • Ticket sales added as a revenue line alongside rides, with staging and production environments defined in Terraform and a canary-guarded release pipeline.

02Kubesense AI · Professional work

Data Explorer and Dashboards

The query builder and dashboards for Kubesense AI's observability platform, built from scratch in React and TypeScript.

  • A visual Builder and a PromQL Code mode for the same query, shaped by requirements from five client teams.
  • Several queries per panel, combined with formulas such as an error rate of B / A.
  • Chart load time cut by 20% on large result sets; core features shipped 50% ahead of schedule.

React · TypeScript · ECharts · PromQL · VictoriaMetrics · ClickHouse

Problem

Engineers needed one place to explore metrics, logs and traces and turn what they found into dashboard panels. Five client teams asked for conflicting things: some wanted a point-and-click builder, others wanted to write PromQL directly.

My role

Software Development Engineer I on the frontend. I built Data Explorer and Dashboards from scratch and worked directly with the client teams on requirements.

Engineering approach

  • A hybrid query experience: a visual Builder for picking a metric, filtering by label and applying functions, and a Code mode for raw PromQL. It satisfied both groups and helped the company secure a proof of concept.
  • Multiple queries per panel, with formulas that combine them into computed series.
  • Dashboards with a drag-and-drop panel grid, dashboard-level variables for filtering, and JSON import and export for sharing and migration.
  • ECharts rendering tuned for large VictoriaMetrics and ClickHouse result sets, which cut chart load time by 20%.

Outcome

  • Core features shipped 50% ahead of schedule.
  • PromQL mistakes only surfaced once the backend rejected a query. That gap later became promql-validator.
  • The public docs describe the product as it is today, which has grown since I left in November 2025.

03Open source · npm

promql-validator

Published on npm

A PromQL editor and validator for observability products: a hand-written, error-tolerant parser, a React editor with autocomplete and quick fixes, Monaco and CodeMirror 6 adapters, and a CLI for CI.

  • Recursive-descent parser checked against a conformance corpus, so it accepts and rejects the same queries Prometheus does.
  • Understands half-typed queries: rate( is reported as incomplete, not as an error.
  • 1,085 tests across 33 suites, plus Playwright end-to-end and automated accessibility checks.

TypeScript · React · Monaco · CodeMirror 6 · Vitest · Playwright

npmGitHub

Problem

At Kubesense, PromQL mistakes only showed up after the backend rejected a query. Most metrics products rebuild the same query box: regex highlighting, a hand-rolled suggestion list, and validation that really means "did the backend accept it".

Parsing and validation

  • Hand-written lexer and recursive-descent parser covering Prometheus 3 syntax, including UTF-8 quoted names, @ start() and @ end(), subqueries and duration expressions.
  • Semantic checks for function and aggregation arity, range-vector arguments, vector matching and conflicting modifiers, with stable diagnostic codes and line and column positions.
  • Reports only the first syntax error, as Prometheus does, and offers quick fixes for common mistakes: rate(up) gets "Add range [5m]".
  • An optional MetricsQL (VictoriaMetrics) dialect and plugins for custom validators.

Editor and developer experience

  • The React component draws highlighting over a native input, so typing, selection, undo, IME and screen readers keep working. It follows the ARIA combobox pattern.
  • Context-aware autocomplete for functions, label names, label values and durations. Async providers are debounced, cancelled when the user keeps typing, cached and protected against out-of-order responses.
  • A zero-dependency core runs in Node, browsers, Deno and workers. The Monaco and CodeMirror adapters share it, so every editor behaves the same way.
  • A CLI lints query files and Prometheus rule files and exits non-zero on an invalid query.

Testing

  • Unit, component, conformance and accessibility tests (1,085 passing), Playwright end-to-end tests against the example app, bundle-size budgets and latency benchmarks.

Outcome

Five MIT-licensed packages on npm: promql-validator and four scoped @promql-validator packages, with documentation and Vite and Next.js examples.

04Personal project

AgroExpert

Working prototype

An AI crop-health assistant for farmers in Nepal who can't reach an agronomist quickly. Crop photos go in; a possible diagnosis comes back with treatment guidance, cited sources, a confidence score and a severity rating.

  • Two LangGraph pipelines: a diagnosis graph for photos and a chat graph for farmers' questions.
  • Each generated claim is checked against retrieved sources before it reaches the farmer.
  • Hybrid dense and sparse retrieval in Qdrant, with reranking.

Python · FastAPI · LangGraph · RAG · Qdrant · PostgreSQL · PaddleOCR · React Native · Expo

Website

Problem

Farmers in Nepal can wait days to see an agronomist, and a wrong guess about a disease or treatment can cost a season's crop. A useful assistant has to show its evidence and be honest about uncertainty.

What I built

A React Native (Expo) app, a FastAPI backend (a modular monolith on PostgreSQL with Celery workers) and a separate AI service. The app talks only to the backend, and the backend is the only caller of the AI service.

AI pipeline

  • Photos are checked first with classical computer vision (format, blur, lighting and screen-photo detection), so unusable images are rejected before any model runs.
  • A vision-language model reasons over all photos together, with the plot's crop, soil type, weather and past diagnoses as context.
  • Retrieval embeds documents with bge-m3, runs hybrid dense and sparse search in Qdrant with rank fusion, then reranks the results.
  • A verification step checks each claim in the recommendation against the retrieved chunks and retries once if the answer is not grounded.
  • Confidence is a weighted score of model certainty, groundedness, context completeness and consistency. Severity is graded none, mild, moderate or severe.
  • OCR (PaddleOCR) reads product labels and scanned PDFs added to the knowledge base.

Safety

  • Input is screened for prompt injection. On output, any step that looks like an invented dosage is removed by a regex check and an LLM check, and the model is told never to name products or doses it cannot ground.
  • Every result carries a localized disclaimer and is framed as preliminary guidance, not an agronomist's verdict.
  • All model calls go through one router with versioned prompts, so models can be swapped per environment without code changes.

Current state

The app, backend and AI service work together end to end on a development stack, and the public website is live. Development runs on CPU with smaller substitute models, so diagnosis quality has not yet been validated with the target models or field data, and the app is not yet publicly released.

Work with me

What I can build for you

freelance & contract

I work with startups and product teams as a freelance full-stack, web and mobile app developer. Remote, from Kathmandu, for clients anywhere.

  • Web applications

    React and Next.js apps that load fast: dashboards, admin and customer portals, and marketing sites built for search.

    Kubesense AI dashboards, MeroGo admin portals

  • Mobile apps

    React Native and Expo apps for Android and iOS, from the first screen to the store release.

    MeroGo passenger and driver apps, Qontentt AI on Play Store

  • Backend and APIs

    Node.js, TypeScript and FastAPI services on PostgreSQL or MongoDB, with REST or GraphQL APIs and real tests.

    MeroGo ticketing service, AgroExpert backend

  • Payments and billing

    Stripe checkout, subscriptions and refunds that never double-charge, with a ledger you can reconcile.

    MeroGo ticketing payments, RevenueCat at Qontentt AI

  • AI features

    OpenAI-powered features, search over your own documents (RAG) and LangGraph workflows that cite their sources.

    AgroExpert, Qontentt AI content assistant

  • MVPs for startups

    Turning a rough idea or PRD into a first version real users can try, then iterating on what they do.

    Led the Qontentt AI MVP frontend

Questions clients ask

Are you available for freelance projects?

Yes. I take on freelance and contract work, and I'm open to full-time roles. Send a short note about what you're building and I'll reply with questions and a plan.

Do you work with clients outside Nepal?

Yes. Since 2024 I've worked remotely with product teams in the US and India.

What time zone do you work in?

Kathmandu, UTC+5:45. My working day overlaps with European business hours and US mornings.

What kind of projects fit best?

Web and mobile apps with a real backend: dashboards, booking or ticketing flows, payments and subscriptions, and AI features on top of your own data.

Which technologies do you use?

TypeScript, React, Next.js, React Native and Expo on the front end; Node.js, FastAPI, PostgreSQL, MongoDB and Redis on the back end; Stripe for payments; OpenAI, RAG and LangGraph for AI features.

Skills

What I reach for

Languages
TypeScript, JavaScript, Python, SQL
Frontend & Mobile
React, Next.js, React Native, Expo, Redux, Tailwind CSS, ECharts
Backend & APIs
Node.js, Express, FastAPI, REST, GraphQL, Stripe, RevenueCat
Data & Infrastructure
PostgreSQL, MongoDB, Redis, Qdrant, Firebase, Docker, Google Cloud Run, Terraform, AWS
AI & Observability
OpenAI API, RAG, LangGraph, PromQL, VictoriaMetrics, ClickHouse
Testing & Delivery
Jest, Vitest, Playwright, Testcontainers, fast-check, k6, CI/CD, PRDs & requirements

Education

Background

Degree

Amrita Vishwa Vidyapeetham

2021 - 2025

B.Tech in Computer Science and Engineering · Kollam, Kerala, India

Chancellor Scholarship (full tuition waiver)

Contact

Say hello

I reply quickly

Open to full-time remote roles, contract work and freelance projects.