Backend and ML engineer / VIT Bhopal + IIT Madras

Parth
Jain.

Most students build for the happy path. I build for edge cases, race conditions, and partial failures.

I build services that stay correct under concurrency and keep running when dependencies fail, and ML systems whose results I can defend. Below is proof, not a list of technologies.

ClientAPIRedis cachedown? use in-memoryCelery queuedown? run synchronouslyPostgreSQLlocal? use SQLitecompare-and-set bookingidempotency keys
How my services are built: every dependency has a fallback.
0.75
Kaggle MAP@3 (LLM logit scoring beat fine-tuning)
0
Double-bookings in concurrent race tests
50%
Less RAM on my Hugging Face matcher
53
REST endpoints in my live Placement Portal

01

Shipped, and running right now

Three full-stack systems with public demos. Each one is built around a failure I wanted to survive, not just a feature list.

01

Placement Portal Application (PPA V5)

Flask / Vue 3 / SQLAlchemy / PostgreSQL / SQLite / Redis / Celery / JWT / Hugging Face / Fly.io / Railway

A real campus problem, not a tutorial CRUD app: students apply to placement drives, recruiters manage drives and candidates, admins oversee everything. Built end to end: schema, API, SPA, caching, async queue and deployment.

53
REST endpoints, documented in api.yaml
12
database tables
14
single-page-app screens
  1. Vue 3 SPA→
  2. Flask API→
  3. Redis cache→
  4. Celery workers→
  5. Postgres / SQLite

Every hop has a fallback: Redis to memory, Celery to a synchronous run, Postgres to SQLite.

  • 53 REST endpoints specified in api.yaml, with Swagger and ReDoc docs, pagination and rate limiting (5/min through Redis).
  • Stateful JWT with a JTI blacklist, so logout really revokes tokens (not just stateless JWTs).
  • Graceful fallbacks by design: Redis to in-memory, Celery to synchronous run, Postgres to SQLite, upload volume to local folder.
More detail
  • 12-table schema with state machines for drives; ATS resume scoring through a Hugging Face model; 14-page Vue 3 SPA with PWA support.

Look at Open the live app (1-click demo login) and read the README's architecture-decisions table, which pairs every tool with the fallback built for it.

02

Healthcare Appointment & Follow-up Manager

FastAPI / SQLAlchemy 2.0 / Alembic / PostgreSQL / SQLite / Redis / Celery + Beat / Google Gemini / Docker Compose / HTMX / Jinja2 / pytest

Patients book slots and submit symptoms, doctors record visits, admins manage schedules and leave. The system is designed to stay correct under simultaneous bookings and to keep working when the LLM, email, calendar or Redis is down.

10 min
slot hold before it expires
5
email retries before dead-letter
0
double-bookings in race tests
  1. Patient picks slot→
  2. Hold (compare-and-set)→
  3. Idempotent confirm→
  4. Outbox + Celery→
  5. Email and Calendar

Gemini triage and summaries always have a deterministic fallback, so the flow never breaks.

  • Concurrency-safe booking: 10-minute slot holds, compare-and-set state changes and idempotency keys, proven by threaded race-condition tests.
  • Gemini pre-visit urgency triage and patient-friendly post-visit summary behind a pluggable provider, with every prompt, output and latency stored for audit; deterministic fallback on failure.
  • Email outbox with retries and dead-letter handling; Celery Beat for hold expiry, reminders and nightly slot generation; leave cancels affected appointments.
More detail
  • Zero-infrastructure quickstart (SQLite, tasks run inline) plus a full Docker Compose topology; DESIGN.md and ARCHITECTURE.md document the hard problems.

Look at The live demo, DESIGN.md (the four hard problems in under 800 words) and the pytest suite covering races, idempotency, RBAC, LLM fallback and the outbox.

03

InventoryPro: Inventory Management SaaS

Next.js 16 / TypeScript / Prisma 6 / PostgreSQL (Neon) / jose JWT / Recharts / Vercel

Replaces fragmented spreadsheets and bloated enterprise software with one app that gives real-time stock visibility and automatic reconciliation for sales, purchasing and manufacturing.

3
order pipelines: sales, purchase, manufacturing
4
stages in a sales order
PWA
installable, with Edge-verified auth
  1. Quotation→
  2. Packing→
  3. Dispatch→
  4. Stock deducted→
  5. Completed

Stock is never edited by hand. It changes only when an order moves to a valid state.

  • Atomic inventory state engine: stock is never edited by hand; a sales order deducts on Dispatch, a purchase order adds on Completion, a manufacturing batch consumes raw materials and adds output.
  • Edge-first auth: JWT verified with jose in Next.js Edge Middleware, so unauthenticated users never reach the server.
  • Recharts dashboards (stock levels, top products, sales status), CSV export, print views, installable PWA, keyboard shortcuts and flicker-free dark mode.
More detail
  • Custom vanilla-CSS design system and a master-detail UI instead of a component library.

Look at Live demo: try /dashboard logged out (instant redirect), advance a sales order to Dispatch and watch stock deduct, and install it as a PWA.

02

Machine learning, with the experiment log

I keep the negative results. They are usually the most useful part: fine-tuning that lost to zero-shot, oversampling that hurt, a metric that lied.

01

Smart MCQ Solver

PyTorch / Hugging Face / Qwen-7B/14B / RoBERTa / DeBERTa / BiLSTM / LightGBM / XGBoost / Weights & Biases / Streamlit

IITM Deep Learning & GenAI project: given a question and five options, return the three most likely answers in order. Documented across 16 approaches, 25+ notebooks and 10 W&B runs.

0.75084
Kaggle private MAP@3
2,000
training rows
16
approaches tried

Public leaderboard score along the way

  • Naive guess (A B C)
    0.270
  • DeBERTa, fine-tuned
    0.352
  • NLI DeBERTa, zero-shot
    0.494
  • Qwen-7B, zero-shot logits
    0.626
  • Qwen-7B, 5-shot logits
    0.705
  • Final ensemble
    0.751

| Cutoff 0.73

  • Honest experiment log: score journey from 0.270 (naive) to 0.75084 private / 0.74937 public, every score from the real leaderboard.
  • Key finding: fine-tuned DeBERTa scored worse (0.352) than zero-shot (0.494); LLM logit scoring with Qwen-7B jumped to 0.705.
  • Final system: three deep models (EmbeddingNN, RoBERTa, BiLSTM with attention pooling) plus gradient-boosted meta-features crossing the 0.73 cutoff.
More detail
  • Also includes a 683-line technical report and a Streamlit app for Hugging Face Spaces.

Look at The score-timeline chart and the negative results (why fine-tuning failed), which show scientific method, not just a final number.

02

FinSight AI: Multi-Agent Financial RAG

LangGraph / LangChain / FAISS / BM25 / Cross-encoder / RAGAS / Groq / FastAPI / SQLite

Answers questions over annual reports and filings with evidence, and checks its own answers for hallucination before replying.

  1. Question→
  2. Input guard→
  3. BM25 + FAISS→
  4. Rerank→
  5. Researcher, Responder, Critic

An output guard checks for hallucination before the answer is returned.

  • Three LangGraph agents: Researcher retrieves, Responder writes, Critic validates against hallucination.
  • Hybrid retrieval: BM25 plus FAISS dense search fused with Reciprocal Rank Fusion, then cross-encoder reranking.
  • Input and output guardrails (prompt injection, PII, hallucination scoring, regulatory disclaimer) and RAGAS-based quality evaluation.
More detail
  • Multi-LLM fallback (Groq, then OpenAI, then Ollama), SQLite conversation memory, FastAPI service.

Look at The architecture diagram and the design-decisions section written as an interview guide.

03

Toxic Comment Classification (Kaggle)

Python / scikit-learn / TF-IDF / LinearSVC / Stacking / XGBoost / LightGBM

A course assignment turned into a five-milestone research process on 197,999 comments with a 20.9:1 class imbalance and a 'no transformers, no deep learning' constraint.

0.826
Macro F1
Top ~150
of 600+ Kaggle teams
20.9 : 1
class imbalance

Macro F1 by milestone

  • Logistic Regression baseline
    0.777
  • LinearSVC
    0.816
  • Stacking ensemble
    0.823
  • Final push
    0.826
  • 60,054-dimension feature space: word 1-3 grams and character 3-5 grams TF-IDF plus 54 engagement, demographic and text-statistics features.
  • Winning model: stacking of four calibrated linear models with a Logistic Regression meta-learner.
  • Macro F1 progression 0.777 to 0.826 across five milestones; LinearSVC gave the biggest single jump.
More detail
  • Tested and rejected SMOTE-style oversampling in favour of class-weight optimisation.

Look at The milestone score progression and the table of all models tested.

04

Solari Demand Forecasting Engine

LightGBM / XGBoost / CatBoost / FastAPI / Pydantic / GitHub Actions / pandas

A Rossmann-style Kaggle problem reframed as a logistics company: over-forecasting wastes perishables, under-forecasting causes stockouts. Adapts the winning approach of Gert Jacobusse.

970K
training rows
1,115
hubs forecast
60+
engineered time features
  • 970,379 training rows and 60+ engineered temporal features.
  • Strategy-pattern forecasters feeding a weighted harmonic-mean ensemble, calibrated with a year-over-year anchor (80/20 blend).
  • FastAPI serving with /predict, /predict/batch and /hubs, typed Pydantic schemas.
More detail
  • Production hygiene: modular src/ package, unit tests and a GitHub Actions CI matrix (Python 3.10 and 3.11).

Look at Software architecture around the model (CI, API, tests), which most Kaggle notebooks lack.

03

Experience

20 May 2026 – 31 Jul 2026

AI/ML & Software Engineering Intern

MPOnline Limited (Govt. of MP & TCS Joint Venture) / Internship

  • Completed two internship tracks (AI/ML, and Advanced Software Engineering & Development), each with a certificate of completion, covering full-stack development, MVC architecture, REST APIs, relational database design, SQL and ER modelling.
  • Built the capstone Smart Retail & Customer Intelligence Platform: OpenCV face recognition, product image classification, sentiment analysis and an FAQ chatbot behind a FastAPI gateway with a Streamlit dashboard.
  • Worked through end-to-end ML pipelines: feature engineering, model evaluation and neural networks for NLP tasks.

Python / FastAPI / Streamlit / OpenCV / scikit-learn / REST APIs / SQL

2025 – Present

Open Source Contributor

Hugging Face / Open Source

  • Built and deployed a two-tier Resume Analyzer & Matcher on Hugging Face Spaces: exact O(N) taxonomy mapping with FlashText, plus semantic gap-filling with KeyBERT and BGE-small embeddings.
  • Cut inference memory by 50% by sharing one SentenceTransformer instance between the embedding encoder and KeyBERT.
  • Implemented hybrid retrieval scoring: asymmetric semantic similarity with BGE query prefixes, two-tier skill coverage with BFS cycle protection, and heuristic quality scoring.
  • Deployed a custom BERT-based NER tool and exposed model endpoints as REST APIs.

Python / Hugging Face Spaces / Sentence Transformers / KeyBERT / FlashText / Gradio

04

Everything else I have built

The index from GitHub, 36 public repositories in total. Click a card to open the code.

Open GitHub

05

What I work with

Grouped by the job it does. Everything here is used in a repository above.

Reach for first

Python / FastAPI / PostgreSQL / Redis / Celery / PyTorch / Hugging Face / Docker

01

Languages

PythonC++CJavaTypeScriptJavaScriptSQLHTML/CSS

02

Backend & APIs

FastAPIFlaskNext.jsVue.jsREST APIsSQLAlchemyPrisma ORMMVCLayered architectureBusiness logic

03

Auth & Security

JWTAuthenticationAuthorizationRBACToken blacklistingCSRF protection

04

Databases & Data Modeling

PostgreSQLSQLiteRelational database designSQL queriesSchema designAlembic migrations

05

Performance & Distributed Processing

RedisCachingCeleryBackground jobsTask queuesAsynchronous processingConcurrency handlingIdempotencyEvent-driven architectureWebSockets

06

Data Engineering & ML

Data preprocessingData transformationFeature engineeringTF-IDFScikit-learnPyTorchLightGBMXGBoostNLPEnsemble learningModel evaluationCross-validation

07

AI / NLP

Hugging Face (Transformers, Spaces)Sentence-TransformersKeyBERTFlashTextSemantic matchingGradioPrompt engineering

08

AI / RAG

LangChainLangGraphVector databases (FAISS)Document chunkingHybrid retrieval (BM25 + dense)Retrieval evaluation (RAGAS)LLM integration

09

Tools & Platforms

GitDockerLinuxGitHub ActionsWeights & BiasesHugging Face SpacesVercelRailwayFly.ioRenderAWS (Educate)Supabase

10

Core CS

Data structures & algorithmsOOPDBMSOperating systemsComputer networksSystem design (HLD/LLD)

06

Education and credentials

B.Tech

Vellore Institute of Technology, Bhopal

B.Tech, Computer Science and Engineering

2023 – Expected Jun 2027 / Full-time, on campus

Core CS: DSA, DBMS, Operating Systems, Computer Architecture, Machine Learning, Deep Learning, Big Data Management.

BS

Indian Institute of Technology Madras

BS, Data Science and Applications (Online)

May 2023 – Expected May 2027 / Part-time, online

Statistics, machine learning, programming and application development, with workshops on Hugging Face, ML in production and LangChain.

IIT Madras BS: progress so far

  1. Foundational Level — Programming and Data Science

    May 2024

  2. Diploma in Programming

    May 2026

  3. Diploma in Data Science

    Sep 2026

  4. Advanced Certificate — Programming and Application Development

  5. Advanced Certificate — Machine Learning and Data Science

Coding profiles

GeeksforGeeks

950+

day coding streak

LeetCode

200+

problems solved

View profile

HackerRank

5 star

Python, Gold level

View profile

Certificates

IIT Madras

  • Diploma in Data Science

    IIT Madras

    Sep 2026

  • Diploma in Programming

    IIT Madras

    May 2026

  • Advanced Certificate: Machine Learning and Data Science

    IIT Madras

  • Advanced Certificate: Programming and Application Development

    IIT Madras

  • Foundational Level: Programming and Data Science

    IIT Madras

    May 2024

Workshops

  • Introduction to LangChain: Chains, Retrieval and RAG

    IIT Madras

    Oct 2026

  • Getting Started with ML in Production

    IIT Madras

    Oct 2026

  • Introduction to HuggingFace and Hugging Face Spaces

    IIT Madras

    Jan 2026

  • NLP with Transformers using Hugging Face

    IIT Madras

    Jan 2025

Courses

  • Internship in AI/ML

    MPOnline Limited (Govt. of MP & TCS JV)

    Jul 2026

  • Advanced Software Engineering and Development Internship

    MPOnline Limited (Govt. of MP & TCS JV)

    Jul 2026

  • Supervised Machine Learning: Regression and Classification

    DeepLearning.AI & Stanford Online (Andrew Ng)

    Jul 2026

  • Introduction to FastAPI Framework

    Duke University

    Nov 2024

  • Introduction to Machine Learning

    IIT Madras

    Apr 2025

  • Marketing Analytics

    IIT Kharagpur

    Apr 2026

  • The Bits and Bytes of Computer Networking

    Google

  • Introduction to Java

    SoloLearn

    Mar 2026

  • Java Intermediate

    SoloLearn

    Mar 2026

  • Introduction to Cloud Computing

    Course certificate

    Jul 2026

HackerRank

  • SQL (Advanced)

    HackerRank

    Jul 2026

  • SQL (Intermediate)

    HackerRank

    Jul 2026

  • Python (Basic)

    HackerRank

  • Problem Solving (Intermediate)

    HackerRank

  • Problem Solving (Basic)

    HackerRank

  • Software Engineer Intern

    HackerRank

Badges

07

How I work

I am a final-year CS student (IIT Madras + VIT) bridging advanced Machine Learning with production-grade backend systems. I do not just write CRUD APIs; I engineer resilient, scalable architectures that behave correctly under real-world constraints.

Architecture and distributed systems

I build decoupled, event-driven backends using a modular monolith approach. This avoids premature microservices complexity while maintaining strict separation of concerns. By leveraging asynchronous message brokers, I ensure heavy background processing never blocks the core user flow.

Concurrency and data integrity

I design compare-and-set state machines to handle high concurrency. They prevent double-bookings and remove check-then-act race conditions, keeping state changes atomic and consistent across concurrent requests without relying on heavy database locks.

Resilience

I build applications that survive partial failures through graceful degradation. If an async broker or cache drops, rate limiters and queues fall back to synchronous or in-memory execution so the core flow keeps working instead of failing. I also build stateful JWT authentication with JTI blacklisting for immediate token revocation.

Open source and applied AI (Hugging Face)

I contributed a semantic matching engine to Hugging Face Spaces. I cut inference memory by 50% with a shared global SentenceTransformer instance, and built a multi-tier NLP pipeline (FlashText taxonomy, then KeyBERT, then BGE embeddings) to replace fragile regex parsers.

Rigorous ML experimentation

I treat ML as a science, not a script. I reached 0.75 MAP@3 on a constrained 2,000-row dataset by showing that zero-shot logit extraction outperforms fine-tuning, which overfit badly, and I deliberately rejected SMOTE for high-dimensional sparse matrices.

Backend and infra
Python, FastAPI, Flask, PostgreSQL, Redis, Celery, SQLAlchemy, Docker, AWS, Fly.io, GitHub Actions
ML and AI
PyTorch, Hugging Face, Scikit-learn, LightGBM, Sentence-Transformers, WandB, TF-IDF
Tools
Linux, Git, Jupyter, VS Code

I am looking for SDE, backend, or ML-focused engineering roles at companies that value engineering rigor, system design, and production-ready code. If you are building systems that need to scale, handle concurrency, or deploy ML models reliably, let us connect.

08

Hiring for backend or ML? Let's talk.

Parth Jain / Bhopal and Chennai, India / Built with Next.js and Tailwind