Bikash Kushwaha standing under an arch of orange flowers

You don't need another model that only works in a notebook. You need one that still works on Monday morning, with real data and real users.

Tell me what you're building →
5× AWS CertifiedML Engineer · Solutions Architect · Data Engineer
CurrentlyAI Engineering at Asterdio · ML at FlyRank AI

I build AI that holds up after launch

The model scored well in the notebook.

The demo impressed the room.

Then real data arrived.

That's not a model problem. It's an engineering problem. And it's the part I'm best at: walk-forward backtests, checkers that catch wrong answers before users do, and releases that roll themselves back when the numbers slip.

For teams that need AI they can measure, trust with private data, and roll back on a bad day. Built in Kathmandu, tested on real Nepali problems.

Selected work

Power grids, air quality, monsoon rain, earthquakes and banking. Every project is measured against a baseline, not a feeling.

01Forecasting · MLOps

Bijuli: Nepal grid demand

Day-ahead hourly electricity forecasts with a model registry, shadow rollout, automatic rollback and drift monitoring.

3.27% MAPEover 23,184 walk-forward forecasts, vs 4.68% for the best naive baseline
View on GitHub →
02Forecasting · Serverless

Hawa: Kathmandu air

Hourly next-day PM2.5 forecasts with alerts in English and Nepali.

40% fewer misseson bad-air days, and 15% lower error than the best naive forecast
View on GitHub →
03Lakehouse · GenAI

Barsha + Sodhnus: monsoon data

A serverless lakehouse over 67,215 station-days of rainfall (1976–2025), with a Nepali/English Bedrock assistant on top, for about $1.20 a month.

97% caughtwrong answers stopped by a deterministic checker in a 350-case adversarial test
View on GitHub →
04RAG · Security

Gopya: secure company RAG

Role-based access enforced inside the vector query, with PII redaction and prompt-injection guards.

0 leaks / 45 attacksand the right document in the top 4 for all 69 test questions
View on GitHub →
05Cloud architecture

Surakshit: earthquake check-in

A multi-region, queue-first service for the ten minutes after a quake, with alert thresholds backtested on 88 earthquakes.

0% lostcheck-ins in a 15-million-user simulation, vs 45% for the naive design
View on GitHub →
06Banking AI · Asterdio

A privacy shield for banks

Three working prototypes for Nepali banks (KYC updates, AML alert triage, credit review) behind a shield that blocks any upload failing a PII leak test.

8 identifiers tokenisedwith maker-checker approval and a full audit trail
Client work · ask me about it →

Why work with me

01

I measure against a baseline.

If a model can't beat a simple forecast, I say so. My grid model scored 3.27% MAPE against the best baseline's 4.68%, over 23,184 forecasts, not one lucky week.

02

I design for the bad day.

Shadow rollouts and automatic rollback cut worst-week error from 9.8% to 7.45%. My earthquake design lost 0% of check-ins at 15 million users.

03

I treat data like it belongs to someone.

Access control inside the vector query, PII redaction, injection guards: 0 leaks across 45 attacks. For banks, a shield that keeps customer data in the country.

The numbers, not the adjectives

0MAPE on day-ahead Nepal grid demand
0of wrong LLM answers caught in adversarial testing
0station-days of monsoon rainfall in one lakehouse
0less manual provisioning time with Terraform

Experience

2024 — now

Nov 2026 – Now

Machine Learning Intern

Weekly ML assignments and a capstone in prompt engineering, AI agents and applied ML, focused on how search engines and AI features read and rank websites.

FlyRank AI
Nov 2026 – Now

Independent Contractor, AI Training

Rubric-based AI training data for a leading AI research lab. Member of Mercor's Machine Learning Engineer Talent Network.

Mercor
Sep 2026 – Now

AI Software Engineering Intern

AI tools for the Breqo app (prompt optimizer, prompt generator, converters) in a Python backend, and three Manus AI banking prototypes with a data-sovereign Privacy Shield.

Asterdio
Jul – Oct 2026

Machine Learning Engineer Apprentice

The full ML lifecycle on AWS, ending in a fraud-detection capstone with S3, Glue DataBrew and SageMaker Studio.

TekBay Digital
Apr – Jun 2026

Solutions Architect

Terraform infrastructure as code that cut manual provisioning time by 40%, least-privilege IAM, and migration runbooks the team still uses.

Adex International
Oct – Dec 2024

Data Engineer

Automated Python data-cleaning and analysis workflows, and charts teams used to make decisions.

DataCamp
Bikash Kushwaha sitting on a sofa in a Kathmandu café
Kathmandu

About me

From data and cloud to AI that ships

I started with data engineering and cloud architecture, so I build machine learning the way an architect builds systems: measured, secured and ready to roll back.

Today I build AI tools at Asterdio, including privacy-first prototypes for KYC, AML and credit review in Nepali banks, and I'm an ML intern at FlyRank AI. Outside work, you'll find me on the cricket pitch or trying the newest tools.

StudyingBSc (Hons) Computer Science with AI, Birmingham City University · 2024–2028
LanguagesEnglish, Nepali, Maithili
Bikash Kushwaha in front of Boudhanath Stupa, Kathmandu
Boudhanath, Kathmandu

Built in Kathmandu. Tested against the worst week of the year.

Every project here starts with a Nepali problem: the power grid, the air we breathe, the monsoon, the next earthquake, and the banks that hold our data.

Contact

Tell me what you're building.

Open to AI and ML engineering roles, internships and collaborations. I reply to every message.

Email me →

kushwahabikash9580@gmail.com