AI Full-Stack Engineer

Hello, I’m Durga.

I build software, explore intelligence, and question what comes next.

Scroll
Stage 1 of 4, Human: a studio portrait of Durga in a black T-shirt against a dark blue backdrop.Stage 2 of 4, Augment: the same portrait wearing smart glasses with a faint cyan display.Stage 3 of 4, Machine: one side of the face replaced by silver plating with cyan light seams.Stage 4 of 4, Beyond: half the face is exposed cyborg metal with a glowing red eye.

If intelligence is no longer exclusively human, what will define us?

  1. HUMAN
  2. AUGMENT
  3. MACHINE
  4. BEYOND

Things I’ve built

Exhibit 01 / 05

Shipped

Immigreat AI

AI Decision Analyzer

A subscription SaaS for asking natural-language questions across 4,600+ EB-2 NIW decisions. Answers come with statistical insights and source-linked citations.

  • Developed a subscription-based AAO Decision Analyzer for natural-language queries, statistical insights and source-linked citations across 4,600+ EB-2 NIW decisions.
  • Engineered a FastAPI AI agent backend with PostgreSQL/pgvector semantic retrieval and metadata filtering, to support evidence-grounded immigration analysis.
  • Delivered a guided question explorer, visual analytics and tiered citation access, with subscription-based question allowances. EB-2 NIW is the first live analyzer.
  • Python
  • FastAPI
  • PostgreSQL
  • pgvector
  • Next.js
  • TypeScript

Exhibit 02 / 05

Shipped

Immigreat AI

GreenCaard LMS & Storefront

The learning side of GreenCaard: a public course storefront and an LMS across 4 immigration categories.

  • Built 2 integrated applications, a public course storefront and an LMS combining student learning and administration, across 4 immigration categories.
  • Implemented Stripe payments, discounts, referrals, refunds and Clerk authentication; secured video lessons with DRM, tracked viewing progress and delivered course assets through Cloudflare R2.
  • Established 3-environment CI/CD with GitHub Actions, Playwright and Vercel; load-tested 1,500 concurrent users using k6 in a production-like environment.
  • Improved reliability and security with Upstash rate limiting, response caching, Zod input validation, CSRF protection, Sentry tracing and structured logging.
  • Next.js
  • TypeScript
  • PostgreSQL
  • Supabase
  • Clerk
  • Stripe
  • Cloudflare R2
  • Cloudflare Workers
  • Upstash Redis
  • Vercel
  • Playwright
  • Sentry
  • k6

Exhibit 03 / 05

Shipped

Immigreat AI

AAO Decision Retrieval Pipeline

Ingestion and retrieval for 14,000+ AAO decisions across immigration categories, with the EB-2 NIW Analyzer currently available to users.

  • Engineered ingestion and retrieval for 14,000+ AAO decisions across immigration categories, with the EB-2 NIW Analyzer currently available to users.
  • Automated rate-limited collection and scheduled processing with Apache Airflow on AWS EC2; extracted structured decision metadata using Anthropic Claude Haiku.
  • Generated Voyage AI embeddings to support semantic retrieval and metadata-filtered discovery across decisions.
  • Python
  • Apache Airflow
  • AWS EC2
  • Anthropic Claude Haiku
  • Voyage AI
  • Web scraping
  • RAG

Exhibit 04 / 05

Prototype

Personal project

Baatein AI

A Hinglish voice companion that remembers you, with your permission. It speaks the mix of Hindi and English people actually use, follows up across calls, and lets you correct or erase anything.

A week with Ram Babu (4:46, sound on): Baatein calls on a demo account across three days of one week, with Hindi and English captions on screen. Ram's lines are scripted; Baatein's replies come from the live agent.
Conversation scenarios
39 / 39
Real model, twice in a row
Right memory first
29 / 31
English
Right memory first
28 / 31
Hindi
Right memory first
24 / 31
Hinglish
Prompt read from cache
98%
From the second reply on
Branch coverage
100%
Strict type checks

Hover or focus a component to trace its connections. Select it to read the note.

  • Python
  • LiveKit Agents
  • Claude Haiku
  • Sarvam
  • SQLite
  • sqlite-vec
  • FTS5
  • Pinecone
  • FastAPI
  • Railway
Read case study

Exhibit 05 / 05

Prototype

Personal project

Lexi

A legal precedent research agent over 56 Indian court judgments, with hybrid retrieval, neural reranking and a 25-query evaluation suite: 0.92 routing accuracy, 0.81 MRR and no fabricated case citations in the test set.

Routing accuracy
0.92
23 of 25 queries
MRR
0.81
Recall@20
0.77
Adverse recall
0.56
Up from 0.34
Faithfulness
3.88 / 5
LLM judge, single sample
Invented documents
0
In the test set

Hover or focus a component to trace its connections. Select it to read the note.

  • Python
  • LangGraph
  • Gemini
  • Pinecone
  • BM25
  • bge-reranker-v2-m3
  • Streamlit
  • LangSmith
  • Railway
Read case study

The future isn’t written yet.

Let’s build something worth exploring.

Explore my workGet in touch

From the blog

All posts