Shubhdeep Das

AI Systems Engineer · Full Stack Developer · Applied Machine Intelligence

[ DOC-ID: TPM-2024-01 ]

Abstract

I build full-stack AI systems that turn messy real-world data — resumes, interviews, speech, video, and medical conversations — into structured, traceable, and useful intelligence. My work sits at the intersection of full-stack engineering and applied AI research, focusing on explainable systems, transcription, resume intelligence, video intelligence, and healthcare AI prototypes.
View Selected SystemsRead Research NotesDownload Resume
[ SEC. 02 ]

The Product Is Clarity

Not the model.Not the score.Not the dashboard.

A system can give an answer and still leave the user confused.

The real work is turning that answer into something people can understand, trust, and act on.

“The purpose of information is not knowledge. It is being able to take the right action.”
Peter Drucker
[ SEC. 03 ]

Projects

SYS-001Active

Resume–JD Intelligence Engine

Research QHow to deterministically align unstructured career histories with specific job requirements?
InputPDF Resumes, Job Description Text
ProcessEntity Extraction, Semantic Matching, Explainable Scoring
OutputStructured Candidate Profile, Fit Analysis Matrix
PythonFastAPILLMs
SYS-002Deployed

Real-Time Multilingual Transcription Backend

Research QHow to minimize latency while maintaining accuracy in live multilingual streams?
InputLive Audio Streams (WebSockets)
ProcessVAD, Stream Buffering, Whisper/Deepgram Inference
OutputDiarized, Translated Text Streams
Node.jsWhisperWebSockets
SYS-003Prototype

Brainrot: Local AI Video Intelligence

Research QCan we achieve high-fidelity video understanding purely on local edge hardware?
InputRaw Video Files
ProcessScene Detection, Frame Extraction, Local Vision Models
OutputSemantic Timeline, Searchable Index
PySceneDetectFFmpegLocal LLMs
SYS-004Research

Doctor–Patient Symptom Graph System

Research QHow to reliably extract medical knowledge graphs from noisy conversational audio?
InputClinical Consultation Transcripts
ProcessNLP Pipeline, SNOMED CT Mapping, Graph Construction
OutputInteractive Symptom Graph, Clinical Notes
[ SEC. 04 ]

Methods & Tools

LAYER 01

Interface Layer

React, Next.js, Nuxt.js, Tailwind CSS

LAYER 02

API Layer

FastAPI, Spring Boot, Node.js

LAYER 03

Intelligence Layer

LLMs, NLP, Whisper, Deepgram

LAYER 04

Media Layer

FFmpeg, PySceneDetect

LAYER 05

Data Layer

PostgreSQL, MySQL, MongoDB, Firebase, SQL Server

LAYER 06

Infrastructure Layer

AWS, Docker, Kubernetes, GitHub Actions

[ SEC. 05 ]

Notes

NOTE-001When Rare Attacks Matter MostProject Showcase / Draft / 14 min
NOTE-002Why Naive RAG Fails When the Task Is Reasoning, Not RetrievalPaper Implementation / Draft / 12 min