iframe-proxy | Sunbelt Computer Software

NbyulaAI engineer

Dec 2024 – now

10,000students use the document system I built
−99%spam after my moderation model shipped
9 in 10admission results logged from email automatically, with no wrong entries
₹80Kin sales from a checkout I rebuilt

See how I built them

Don’t take my word for it. Ask my AI anything about my work, and it answers from my real resume.

3 months to hire an AI engineer

+ 6 weeks of demos that never reach real users

+ 4 weeks building evals nobody trusts

+ every week hand-tuning prompts

= 6+ months and still no AI in production

Or talk to someone who’s already shipped it.
Book a free 15-min call

2026 Featured in ETV Bharat

Hey, Traveler from the internet

I’m Nahfid, an AI engineer at Nbyula. I’ve shipped models that 10,000 students rely on and published 5 papers.

I learned that most AI never leaves the demo. It wows a meeting, then nobody trusts it with real users.

Prototypes, slide decks, one-off prompts… but does any of it actually run for real users on a Monday morning?

So I build AI by 3 rules:

Start with the eval, not the prompt.
Ship to real users early, then fix what actually breaks.
Automate the boring loop. My prompt agent tests its own rewrites and opens its own PRs.

If you need AI that ships, let’s talk for 15 minutes.

— Nahfid

Experience

Where I’ve shipped

Numbers first. Click any role for how I built it.

Software Development Engineer 1

Nbyula · EdTech, Bangalore

Jan 2026 – now

388Kpast decisions behind my admissions model

9 in 10admission results logged from email automatically, with no wrong entries

1serving layer for every AI agent, with failover

0humans in the prompt loop. My agent opens its own PRs

How I built it

Admissions model that gives a student their real chance at a specific university: LightGBM over 388,000 past decisions, each university scored against its own admit history, validated on held-out intake cycles.
Automated recording of university decisions that used to be read out of inboxes by hand: a classifier filters the mail, an extractor pulls the decision (rules first, model where they miss), and a resolver maps it to the right application.
Consolidated the product’s AI agents onto one serving layer: pooled capacity with automatic failover to a metered API, a shared credential store, and a single async client.
Self-improving prompt agent (GEPA) that mines real conversation traces, drafts rewrites, scores them on an eval set, and opens its own PR when one wins.

SDE Intern, AI & Full-Stack

Nbyula · EdTech, Bangalore

Dec 2024 – Jan 2026

10,000users on my document pipeline

−99%spam after my moderation system shipped

₹80Kproduct-led sales from a rebuilt checkout

3AI agents shipped: WhatsApp, Instagram, browser

How I built it

The product’s first AI agents: a LangGraph + WhatsApp bot that re-engages dormant leads from CRM history, an Instagram DM agent via ManyChat, and a Browser Use agent that files foreign university applications end-to-end.
Document pipeline behind 10,000 users: uploads queue through BullMQ into OpenAI Batch for classification and extraction, an ONNX model on Lambda straightens crooked scans, and unreadable files are sent back before they stall an application.
Content moderation: a fine-tuned BERT classifier with an LLM second opinion.
Rebuilt the SOP and Essay flows around direct Razorpay checkout.
Extended the CRM with lead filtering, call and Google Meet auditing, and an Operations role.

SDE Intern, Full-Stack

CreditMitra · FinTech, Hyderabad

Apr – Sep 2024

0 → 1built creditmitra.in end to end

CI/CDJenkins + Ansible deploys on AWS

1internal HRMS app: attendance, payroll, offer letters

How I built it

Took creditmitra.in from zero to live: a MERN lending platform on AWS, served through S3 + CloudFront with Route 53, with separate access levels for customers, staff and admins.
Set up Jenkins + Ansible so every release ran through a pipeline instead of by hand on the server.
Wired server-side analytics (Microsoft Clarity, Google Tag Manager, Meta Pixel) to track where sign-ups came from and where users dropped off.
Built the team’s internal HR app, an installable PWA for attendance, payroll, offer letters and the staff directory.

Work