Blueprint / D. E. Joseph

Sheet 01 — Subject

Deepika
Elsa Joseph

AI/ML & Data Science Engineer

I build the pipeline layer, not just the model on top of it. Raw signal turned into validated analytical bases. Retrieval that resolves which version of a record is authoritative. Two IEEE systems shipped to completion.

View projects Read the papers
  • 2IEEE papers
    indexed on Xplore
  • 89.3%retrieval accuracy
    up from 61.4%
  • 100%on-device inference,
    no data leaves the machine
Watercolour portrait painted by Deepika Elsa Joseph Watercolour portrait painted by Deepika Elsa Joseph

Idukki, Kerala

the hard part is
upstream of the model

hallucination rate

24.7% → 8.4%

at 94.2% smaller footprint

Subject
D. E. Joseph
Discipline
AI/ML · Data
Sheet
01 of 06
keep reading

Sheet 02 — About & technical skills

Who’s on the drawing

The short version

I’m an Integrated M.Tech Data Science student at SRM IST, and most of my work lives one layer below the model — ingestion, transformation, schema, retrieval correctness.

I’ve built a version-aware index over 15,000+ legal documents, where the hard part was never the LLM — it was designing retrieval that knows which version of the law is the authoritative one. A local-first analysis service that keeps source code on the machine it came from. A primary field study of 147 workers, from the collection instrument through to the statistics. And a transformation layer turning raw industrial sensor streams into a model-ready feature set. Two of them became IEEE papers.

I prefer controlled comparisons over assumptions: identical inputs, identical metrics, every cell of the matrix directly comparable. And I write the thing down afterwards, so someone else can run it again.

— Deepika

Technical skills

Programming

  • Python
  • SQL
  • TypeScript
  • JavaScript

Machine learning

  • PyTorch
  • TensorFlow
  • Scikit-learn
  • XGBoost
  • ANN / ANFIS
  • Cross-validation

AI & LLM systems

  • RAG
  • Fine-tuning
  • FAISS embeddings
  • Chunking strategy
  • Local inference (Ollama)

Data science & statistics

  • pandas
  • NumPy
  • Hypothesis testing
  • ANOVA
  • Chi-square
  • ROC-AUC

Data engineering

  • ETL / ELT design
  • Ingestion & validation
  • Schema design
  • Signal preprocessing (FFT)
  • Streaming delivery
  • PostgreSQL
  • MongoDB

Tools & infrastructure

  • Git
  • GitHub Actions
  • Docker
  • Linux
  • CUDA
  • FastAPI
  • Node / Express

Sheet 03 — Built & shipped

Things I made work

2026

StudySync

Full-stack data platform

  • Designed the MongoDB schema and Express API around the access patterns the progress dashboard actually needed — not modelled first and queried around afterwards.
  • Built the full processing path: input → transformation → external LLM call → structured output parsing → storage → rendering, with unstructured responses normalised into a predictable schema before persistence.
  • React
  • Node / Express
  • MongoDB
  • LLM API

2024

CodeReviewBot

Local-first analysis service

  • LLaMA 3.2 running fully on-device via Ollama — 100% local execution, so source never leaves the machine, addressing the data-residency constraint that blocks cloud tooling in regulated environments.
  • 6 independent analysis modes over one shared backend as separable reusable components, with a streaming pipeline delivering incremental output instead of blocking on completion.
  • FastAPI
  • Python
  • React
  • TypeScript
  • Docker
  • Ollama

SIH 2024

Crop Disease Prediction

IoT → CNN → farmer dashboard

  • Built the ingestion path from IoT image capture through a CNN inference backend over REST to a farmer-facing dashboard, delivering classification within seconds of capture.
  • National Finalist, Smart India Hackathon 2024 — India’s largest national-level hackathon.
  • React
  • CNN
  • OpenCV
  • IoT
  • REST

Sheet 04 — Research & publications

Where the evidence is

IEEE ICDSBS 2026, National University of Singapore — both indexed on IEEE Xplore

Paper 01 · second author

Distilling Domain Knowledge into Smaller LLMs using RAG-Assisted Fine-Tuning

  • Built the ingestion and indexing pipeline for a corpus of 15,000+ chunked legal documents, from raw source parsing through chunking strategy to a queryable FAISS vector index.
  • Solved a data-lineage problem at the index level: temporal-aware retrieval so queries resolve against the authoritative version of a superseded framework — correct framework selection 61.4% → 89.3% across the July 2024 IPC→BNS transition.
  • Ran a controlled 4×4 evaluation matrix across 24 configurations with inputs and metrics held constant, making every cell directly comparable.
  • −66%hallucination rate
    24.7% → 8.4%
  • −94.2%storage / compute
    footprint
  • <2%latency cost
Read on IEEE Xplore

Paper 02 · primary data study

Wheels of Burden: A Statistical Analysis of Economic Pressure, Mental Health, and Road Safety Risk Among Gig Economy Workers in India

  • Designed the collection instrument and the full preprocessing pipeline for a primary dataset of 147 gig-economy workers across two cities, owning the path from raw collection to cleaned analytical base.
  • Applied correlation, ANOVA and chi-square testing to isolate statistically supported risk factors.
  • Translated the findings into a written argument for a non-specialist audience and presented at an international IEEE conference.
  • 147respondents,
    2 cities
  • 3statistical tests
    applied
Read on IEEE Xplore

Sheet 05 — Experience & education

On the record

CSIR-CSIO

Research Collaborator — Industrial Data Systems, under Principal Scientist supervision

Dec 2024 – Dec 2025Tamil Nadu, India

  • Owned the ingestion and transformation layer of an industrial fault-detection system end to end — built the FFT pipeline converting raw vibration sensor streams into a structured, model-ready feature set, designed as one reusable component feeding three downstream consumers rather than three parallel one-off scripts.
  • Ran source-system reconciliation across 8 motor fault categories: generated the synthetic dataset, validated it against real-time motor readings category by category, and documented every divergence between generated data and field behaviour.
  • Benchmarked three model families (ANN, ANFIS, Fuzzy Inference) on identical inputs and evaluation criteria, so the recommendation rested on a controlled comparison rather than assumption.

Education

SRM Institute of Science and Technology

Integrated M.Tech, CSE (Data Science)

Kattankulathur, Chennai · 2022 – May 2027

Five-year integrated programme. Alongside it: a year of external research collaboration and two conference papers.

Achievements

  • Two IEEE papers accepted, presented and indexed on IEEE Xplore — ICDSBS 2026, NUS.
  • Smart India Hackathon 2024 — National Finalist.
  • Work that spans research, industrial R&D and product — published retrieval systems, a year of sensor-data collaboration, and full-stack tools people actually run.

Sheet 06 — Sign-off

Let’s talk pipelines

Open to AI/ML and data engineering roles and internships — especially the kind where the data layer is the hard part. Fastest way to reach me is email.

Emaildeepikaelsajoseph@gmail.com LinkedInin/deepika-elsa-joseph GitHubgithub.com/deepikaelsa PublicationsTwo IEEE papers on Xplore

thanks for reading the whole sheet.