Apr – Jul 2024Canada, Remote
EaseworkAI
Full-Stack & AI Engineer Intern
Automated supplier–buyer procurement email end to end with CrewAI agents, Microsoft Graph, FastAPI and LLMs.
~60%
faster turnaround
agents working real inboxes
Hi, I'mYash. Yash Varma, AI Product Engineer.
the person behind the code
I build full-stack products, applied AI systems and the automations that tie them together, usually from a messy first idea all the way to production.
AI Product Engineer · MSc CS at UCD · 10+ hackathon wins
Chapter 012021
It started with a phone.
In 11th grade I didn't own a computer. So I installed Pydroid on my phone and started writing Python on a screen smaller than my palm.
A notepad. A calculator. A little directory for saving contacts. Command-line programs, typed one line at a time, and every one of them ran.
- a CLI notepad
- a calculator
- a contact saver
# calculator
a = float(input("a: "))
b = float(input("b: "))
op = input("+ - * /: ")
if op == "+":
print(a + b)
print("hello, world")
No fancy setup. Just curiosity.
$ python notepad.py
>
Chapter 022022 – 2024
Then I started building.
In 2022 I started a BE in Artificial Intelligence & Machine Learning at SIES Graduate School of Technology in Mumbai. The first year pulled me in three directions at once.
- 01
Python, properly
Data structures, algorithms, and the theory behind the scripts I'd been writing.
- 02
The web
Learned web development and built my first portfolio.
- 03
Algorithms
My first contest: Top 20 in CodeChef's national Turing Cup, 2023.
which turned into running things
- 2023 – 24
Competitive Programming Coordinator · Arena SIES GST
Curated 50+ problems and ran contests for ~500 participants.
- 2024 – 25
Technical Lead · E-Cell SIES GST
Led full-stack work on internal platforms and mentored 20+ junior developers.
The degree
BE · AI & Machine Learning
SIES GST, University of Mumbai
2022 – 20268.95CGPA, out of 10
Chapter 032023 – 2026
Ideas started becoming products.
Somewhere along the way the code stopped being just for me. Teams, clients, real users, and deadlines I hadn't set myself.
After first year
HFC Enterprises
An unpaid, three-month internship reviewing Java and Python code at a startup run by one of my seniors. It's where I learned how professional work actually gets delivered.
unpaid, and worth it- Jan – Feb 2025San Francisco, Remote
Cartman Labs (Slaps)
Software Development Engineer Intern
Shipped an autonomous X agent on Eliza OS, plus a Web3-authenticated community platform on DigitalOcean and Vercel.
99.5%uptimeschedules & replies on its own - Jun – Jul 2025Mumbai, On-Site
ConnectWise
Engineering Intern
Built a RAG issue-analysis dashboard over 10,000+ Jira tickets, fed by a cron pipeline with incremental updates and retries.
~75%less manual triageoffered a full-time role → chose the MSc - May – Jul 2026UK, Remote
Norxio
Founding Software Engineer
Started as a freelance site build, became B2B payment and FX dashboard flows in Next.js and Nest.js, on Cloudflare Workers.
~90%fewer translation API callsoffered equity. still picked Dublin.
In between, I took on freelance client sites: designed, built, deployed and tuned for performance and SEO.
Full experienceChapter 042023 – 2025
I like building under pressure.
My first hackathon was Hackanova 3.0. In 36 hours we built FactFinder, a BERT-based fake-news detector, and placed 1st runner-up out of ~2,500. After that I kept coming back.
10+
wins, and counting- $ kavach scan ./suspect.bin› sandboxdisposable Docker container› staticfile entropy analysis› dynamicrule-based syscall flags› modelXGBoost · ~184 GB of malware bytes› verdictexplained by LLaMA 3.2, in plain wordsvalidated live on GajShield's own samples1st of 2,000+ teams, GajShield Hack 8
Kavach.ai
- the problem
- Suspicious binaries are risky to open and slow to analyse by hand.
- what we built
- A desktop suite that detonates files in a disposable Docker sandbox, scores them with an XGBoost model trained on ~184 GB of malware, and has LLaMA 3.2 explain the verdict.
- why it matters
- Analysts get a verdict they can read, not just a score.
Electron.js, Python, Flask, Docker, XGBoost, LLaMA 3.2Code - Loan agreementPaddleOCRMistral Saba + FinLLaMa8 languagespredatory clause?1st of ~5,000, Great Bengaluru Hackathon
LoanSaathi
- the problem
- Loan paperwork is dense, and predatory clauses hide in it, often in a language the borrower doesn't read.
- what we built
- In 24 hours: PaddleOCR reads the documents, Mistral Saba 24B and a fine-tuned FinLLaMa flag risky clauses, and the assistant walks through eligibility and EMIs in 8 Indian languages.
- why it matters
- Documents are discarded after every session. Private by design.
FastAPI, Next.js, MongoDB, Supabase, Mistral Saba 24B, FinLLaMa, PaddleOCR, Sarvam AICode - Stock analysissimulated tradesLive stock alertdelivered on WhatsApp via Twilioalerts land on WhatsApp1st of 2,000+, Datathon 24
FinSaathi
- the problem
- Markets and government schemes are hard to navigate if finance isn't your job.
- what we built
- A finance platform for stock analysis, trade simulation and scheme-eligibility discovery, with live stock alerts on WhatsApp via Twilio.
- why it matters
- It meets people in an app they already open every day.
- PFZborderSOSshown to ISRO's ChairmanMost Product Readiness, Hackcelestial 2.0
SeaGuard
- the problem
- Fishing crews need to know where the fish are, and where the maritime border is.
- what we built
- A React Native assistant predicting Potential Fishing Zones at ~97% accuracy from satellite and weather data, with SOS alerts, boundary warnings and multilingual voice.
- why it matters
- Livelihood and safety in one app. Exhibited to the Chairman of ISRO in 2025.
React Native, FastAPI, TensorFlow, Python, Supabase, Sarvam AICode
proof, from the days themselves
1st · The Great Bengaluru Hackathon 
1st · GajShield Hack 8, KJSCE 
1st · Datathon 24, KJSCE 
Best · Hackcelestial 2.0 
3rd · AMD AI Sprint, IIT Bombay 
2nd · Hackanova 3.0, TCET 
ISRO · Conquer Space Exhibition
Also 1st at Cognition SIES GST, 3× Winner of Smart India Hackathon (SIH) Internal (2023, 2024, 2025), Top 12 Finalist at Sarvam BuildIn' Hours (2026), and Top 20 in CodeChef Turing Cup.
All achievementsChapter 052024 – 2026
Curiosity took me into research.
Some problems don't fit in a weekend. These are the ones I kept digging into: vision, speech, language and hyperspectral imaging.
- Manuscript in preparation
S2VNet-CA: Cross Attention Enhanced Subpixel Spectral Variability Network
First author · Final-year thesis
Cross-attention between a spectral-unmixing branch and a spatial branch, for hyperspectral classification when labels are scarce.
- Published · IEEE ICCUBEA 2025
Attention Enhanced Speaker Representation with Contrastive Learning
Co-author
An attention-enhanced speaker recognition pipeline trained with contrastive learning.
- Accepted · Springer ICDSM 2024
Advanced Fake News Detection Using BERT and Ensemble Learning
Co-author
BERT embeddings with an ensemble of classical classifiers: ~95% accuracy on the 72,000-article WELFake dataset.
this one began as FactFinder - Registered software copyright · 2025
Dimension Capture: High-Precision ArUco Computer Vision System
Computer vision · ArUco markers
ArUco-marker measurement from 10 cm to 5 m within 2 mm, at ~90% lower cost than LiDAR.
Chapter 06Sep 2026 →
Now, Dublin.
In September 2026 I moved from Mumbai to Dublin for an MSc in Computer Science at University College Dublin.
University College Dublin (UCD)
MSc in Computer Science (Negotiated Learning) · Sep 2026 – Sep 2027
It's a negotiated-learning programme, so I'm shaping the modules around what I want to get better at:
- Distributed Systems
- Advanced Machine Learning
- Multi-Agent Systems
- Software Engineering
- Big Data Programming
MumbaiDublin~7,600 km later
it's here right now
and what I'm doing with it
Chapter 07Today
Still building.
What I like most is the whole arc: a messy problem, a system that solves it, and the unglamorous part where it runs in production and real people use it.
Right now I'm building
same person, still curious