Selected work,
cased and shipped.

An end-to-end pipeline that turns unstructured, real-world documents — scanned PDFs, photographed forms, multi-page invoices — into clean, validated JSON ready for downstream systems.

How it works: layout-aware chunking → OCR with Tesseract and a PaddleOCR fallback for low-quality scans → vision-LLM extraction → strict JSON-schema validation → human-in-the-loop review queue for the long tail of edge cases.

Why it matters: the team replaced weeks of manual data entry per client onboarding with a system that processes a document in seconds and learns from every correction.

Designed two retrieval architectures side-by-side — a simple RAG for fast lookups and an advanced hybrid RAG for nuanced questions — built on LangChain and LlamaIndex with PgVector and Mongo Vector as the backing stores.

The hybrid pipeline blends dense embedding similarity with BM25 keyword reranking, then runs LLAMA inference on the assembled context. A small evaluation harness compares variants on the same question set so we can ship retrieval changes with confidence rather than vibes.

For high-volume workflows, I added SLM-first routing for narrower tasks and reserved larger models for complex reasoning paths, reducing inference cost by around 30%.

The result was fewer confidently-wrong answers and a measurable drop in user-flagged hallucinations.

GenAI services fall over differently than ordinary web services — long-running streams, provider rate limits, partial responses, and multi-region quota math. I built the deployment + observability layer that keeps ours upright.

CI/CD with GitHub Actions and Jenkins, Docker images per service, Helm charts on Kubernetes, blue/green deploys to AWS ECS, Prometheus + Grafana for metrics, and structured logs that make a stream's life story replayable.

I also built queue-backed execution with Celery workers for long-running generation and report jobs, improving responsiveness for interactive APIs.

The headline win: persistent 429 RESOURCE_EXHAUSTED errors mid-stream were resolved with a custom retry strategy — exponential backoff, token-bucket smoothing, and graceful resumption — that protected uptime without doubling our bill.

Shipping LLM features means inheriting a new attack surface. I led the work that wrapped our GenAI services in a thin, opinionated guardrail layer.

The layer covers prompt-injection detection, PII redaction in front of every model call, structured output validation against allow-lists, and dependency-level vulnerability scanning in CI. The same models we ship are repurposed to flag suspicious patterns — model output as observability signal, not just product output.

Published research extending traditional OCR to extract text from video for indexing and analysis. The system was built in Python with OpenCV2, PyTesseract, and a speech-recognition module.

The pipeline combines video indexing, text segmentation and recognition via the Stroke Width Transform (SWT) + Tesseract, and parallel speech recognition with MoviePy — producing a fully searchable digital transcript of arbitrary video content.

At Arohak Technologies Pvt Ltd, I built and deployed full-stack web applications using FastAPI or Node.js services with React/Next.js frontends.

I implemented JWT-based authentication and authorization, optimized relational and NoSQL data models, and containerized services with Docker for scalable AWS deployment workflows.

This role was where I hardened the engineering fundamentals I now apply to GenAI systems: API design, secure auth, database performance, frontend quality, and production deployment discipline.