IndexWorkAbout PlaygroundContact
Index / About

The engineer
behind the résumé.

Four-plus years of building software, with deep production focus on GenAI, RAG, and full-stack AI products. I am opinionated about evals, allergic to hand-wavy AI demos, and happiest when something I built is quietly running for real users.

Short story.

§ — Background

I trained as a computer scientist (Master's at the University of North Texas) and got hands-on with classical ML and computer vision before LLMs ate the world. That foundation matters — I treat models as components in a system, not magic, and I default to evaluating them like any other piece of code.

Today I'm a Full Stack AI Engineer at Sunus AI, where I own GenAI features end-to-end: from the retrieval architecture and the prompt + tool design, through the FastAPI service that serves them, all the way to the React / Next.js surface a customer actually clicks on.

Before that I was a graduate student assistant supporting CS faculty, doing the unglamorous-but-important work of keeping research infrastructure alive while researching emerging AI/ML techniques on the side. The first paper I co-authored — on Video OCR + the Stroke Width Transform — came out of that period.

My earlier role at Arohak Technologies Pvt Ltd grounded my full-stack foundations: API design, secure auth, database performance, and cloud-oriented deployment workflows.

Experience.

§ — Timeline
Mar 2025 — present
Live

Full Stack AI Engineer

Sunus AI

Lead end-to-end development and deployment of GenAI SaaS applications, leveraging frontier LLMs and integrating them into robust user-facing web platforms with a full-stack AI operations mindset.

  • Built an Intelligent Document Processor combining GenAI + computer vision to automate extraction from complex unstructured documents.
  • Built advanced and lightweight RAG pipelines (LangChain / LlamaIndex) with PgVector and Mongo Vector, reducing hallucinations and improving answer grounding.
  • Reduced inference costs by 30% by routing targeted tasks to SLMs and reserving larger LLMs for higher-complexity reasoning.
  • Implemented asynchronous workload handling with Celery worker pools and queue-based orchestration for long-running AI tasks.
  • Strengthened platform security using OAuth 2.0 enterprise auth, prompt-injection guardrails, PII redaction, and dependency vulnerability scanning.
  • Built observability and cost analytics workflows (Prometheus/Grafana + usage reporting), contributing to a 15% cloud compute spend reduction.
  • Resolved persistent 429 RESOURCE_EXHAUSTED errors in streaming generation with custom retries and exponential backoff.
Jun 2023 — Dec 2024
Graduate

Student Assistant

University of North Texas

Led GenAI-focused research and engineering support while pursuing my Master's degree.

  • Built a LangChain-based RAG pipeline and full-stack research app for secure internal knowledge retrieval.
  • Developed FastAPI inference APIs and retrieval pipelines, improving average response time by 20%.
  • Managed ingestion workflows (chunking, embeddings, indexing) and supported open-source LLM experimentation.
  • Containerized AI services with Docker and automated CI workflows for reproducible research deployments.
Jun 2022 — Dec 2022
Foundation

Full Stack Engineer

Arohak Technologies Pvt Ltd

Built and deployed full-stack web applications across backend APIs, frontend UX, and production infrastructure.

  • Delivered backend services in FastAPI / Node.js with React / Next.js frontends.
  • Implemented secure authentication and authorization with JWT-based patterns.
  • Designed and optimized PostgreSQL, MySQL, and MongoDB data workflows.
  • Containerized services with Docker and supported scalable AWS deployments.

Skills atlas.

§ — Tools / methods

The honest version. Filled dots = years of real, in-production use; dim dots = comfortable but lighter mileage. The headline stack at the top is what I'd reach for on day one.

01
Python — backend & AI glue
Native tongue · FastAPI · Flask · Django · type-checked
5 yrs
02
GenAI & RAG
LangChain · LlamaIndex · PgVector · evals · LLM orchestration
3 yrs
03
React & Next.js
App router · TypeScript · Tailwind · production SaaS surfaces
4 yrs
04
Cloud & containers
AWS · Docker · Kubernetes · CI/CD · multi-cloud deploys
3 yrs
05
Data stores
PostgreSQL + PgVector · MongoDB · Redis · schema design
5 yrs
06
MLOps & observability
GitHub Actions · Jenkins · Prometheus · Grafana · structured logs
2 yrs
A · 08

GenAI & LLMs

RAG (advanced + hybrid)
SLM/LLM routing
Prompt engineering
Vector databases
LLM evals & harnesses
Fine-tuning
Transformers (theory)
B · 07

AI / ML frameworks

LangChain
LlamaIndex
Hugging Face
PyTorch
TensorFlow
scikit-learn
CrewAI / multi-agent workflows
C · 04

Languages

Python
TypeScript
JavaScript
SQL
D · 05

Web / front-end

React
Next.js
Tailwind CSS
HTML & CSS (this site)
Animations / motion
E · 04

APIs / back-end

FastAPI
Flask
Django
Node.js
F · 07

Cloud & DevOps

AWS (ECS · EC2 · Lambda)
Docker
Kubernetes
CI/CD (GH Actions · Jenkins)
Celery async orchestration
Nginx
Git
G · 06

Data stores

PostgreSQL
PgVector
MongoDB · Mongo Vector
Redis
MySQL
H · 05

Trust & safety

PII redaction
Prompt-injection guards
OAuth 2.0 + access control
Schema / output validation
Vuln scans & supply chain
I · 03

Currently learning

Long-context evals
Agentic tool-use
Latency budgeting

Education.

§ — Formal training
2022 — 2024

Master of Science · Computer Science

University of North Texas

Coursework and research across machine learning, computer vision, and software systems. Co-authored published research on Video OCR with the Stroke Width Transform during this period.

How I work.

§ — Principles

01 · Ship the eval first

If we can't measure a change, we shouldn't ship it. I write the evaluation harness before I tune the prompt.

02 · Boring infra wins

Postgres before Pinecone, FastAPI before frameworks-of-the-month. The model can be exotic; everything around it should be calm.

03 · Latency is a feature

Streaming, perceived speed, sensible loading states — UX is part of the AI system, not a wrapper around it.

04 · Guardrails on every call

PII redaction, prompt-injection screens, schema validation. Treat the LLM like an untrusted external service.

05 · The full stack matters

An AI feature only counts when someone clicks a button. I work the whole way down to that button.

06 · Write things down

Decisions, trade-offs, and post-mortems. Future-me and the next engineer will thank present-me.