About Experience Projects The Lab Contact

Building intelligent systems that scale.

I'm an AI & ML Engineer specializing in production LLM backends, RAG pipelines, multi-agent systems, and reinforcement learning. B.Tech Mathematics & Computing student at PEC Chandigarh.

Enter the Lab GitHub LinkedIn
Manas Ranjan Jena

Manas Ranjan Jena

Mathematics & Computing @ PEC Chandigarh

I build production-grade ML systems — from async LLM backends and multi-agent pipelines to real-time RAG routers and CV pipelines. Passionate about the intersection of mathematical theory and robust software engineering.

About Me

Professional Overview

Program Achievement

Summer of Bitcoin 2026

Shortlisted from 60,000 applicants worldwide

Successfully cleared the competitive Week 1 cryptography and protocol development challenges. Contributing to open-source Bitcoin core developer tools and scaling infrastructure (LND/gRPC and Go).

17+
GitHub Repos
3
Hackathons
6
Simulations
2028
Graduating
Targeting Roles
ML Engineering Intern
AI/LLM Backend Intern
Research Intern (RL / GenAI)
MLOps Engineer Intern
Languages & ML Stack
Python Golang C++ Java PyTorch TensorFlow Scikit-learn LangChain Gymnasium Hugging Face
Backend & Infrastructure
FastAPI Celery Docker Redis PostgreSQL MongoDB ChromaDB MLflow
Timeline

Where I've shipped things

2026
Shortlisted

Summer of Bitcoin 2026

Summer of Bitcoin — Open Source

Shortlisted from 60,000 applicants worldwide for the prestigious open-source program. Cleared Week 1 cryptographic and core protocol development challenges successfully. Engaging with Bitcoin protocol development, transaction building, and lightning network architectures.

Bitcoin Core Cryptography Go C++ gRPC
2025 – Present
Freelance

Freelance AI Engineer

Self-Employed

Built and deployed production LLM backends using FastAPI, Celery, and Docker. Designed async task queuing with Celery + Redis for concurrent ML inference workloads. Specialised in lightweight, resource-efficient ML deployments via containerised model serving.

FastAPI Celery Docker Redis LLMs
2025
Hackathon

Sole AI & Backend Engineer — MindHaven

Google GenAI Hackathon 2025

Built a production RAG pipeline with LangChain, ChromaDB, and Gemini featuring hybrid retrieval, multi-turn memory, and Fernet-encrypted user-isolated storage. Implemented JWT auth, rate limiting, Celery task queuing, STT/TTS voice support, async streaming, and a modular FastAPI backend. Designed and ran an LLM evaluation framework comparing Gemini models across empathy, cultural relevance, and safety metrics.

LangChain ChromaDB Gemini FastAPI Celery Fernet STT/TTS
2025
Hackathon

Team Lead & Sole AI/Backend — Samudra Sachet

Smart India Hackathon 2025

Built real-time CV + NLP pipelines for hazard classification and misinformation filtering. Implemented DBSCAN geospatial clustering with PostGIS for live disaster hotspot detection. Designed FastAPI + PostgreSQL + Cloudinary + Celery backend for async media processing and alert generation.

OpenCV NLP DBSCAN PostGIS FastAPI PostgreSQL
Mar 2025
Hackathon

Financial Report Generator

IIT Kanpur — Techkriti GenAI Hackathon

Built an end-to-end LLM pipeline converting raw financial datasets into structured visual reports. Resolved latency by optimising vector store queries and API call chains.

LLMs Vector Store FastAPI MongoDB
Repositories

Things I've built

SwiftPredict

I got tired of copying and pasting preprocessing and model training loops for every new dataset, so I built SwiftPredict. It's a single-command AutoML library that handles imputation, encoding, and training end-to-end. I rewrote v2 to make the CLI deterministic and published it to PyPI.

  • Reduced ML experimentation time by ~90% and boilerplate code by ~95%
  • Published to PyPI — pip install swiftpredict-v2
Python Scikit-learn FastAPI PyPI

MindHaven

My submission for the Google GenAI Hackathon 2025. I built a production-grade RAG pipeline from scratch. It features user-isolated database encryption using Fernet, multi-turn conversation memory, and streaming audio support. I also ran LLM evaluation scripts to benchmark Gemini's empathy metrics.

LangChain ChromaDB Gemini Celery FastAPI

Keiro

An adaptive RAG gateway proxy and query router service. Written in Golang (router proxy and semantic cache gateway) and Python (gRPC intelligence service classifying query complexity). Minimizes LLM usage costs and query latency.

  • Low latency gRPC orchestration between Go gateway and Python services
  • Embedded LRU semantic cache resulting in 0.8ms average query response times on hits
Golang Python gRPC ChromaDB Redis

Multi-Agent Refiner

I designed a multi-agent system where five specialized LangChain agents (Writer, Syntax Fixer, Optimizer, Doc Generator, and Reviewer) peer-review and optimize code. By orchestrating their communication using FastAPI, I observed a 40% improvement in complexity metrics.

LangChain FastAPI MongoDB OpenAI

DCGAN Digit Generator

To deeply understand GAN dynamics, I built a DCGAN from scratch using PyTorch. I ran into mode collapse early on, which I resolved by implementing one-sided label smoothing and mini-batch discrimination. I tracked FID scores over 200 epochs to measure visual quality.

PyTorch GANs MNIST

Samudra Sachet

I was the Team Lead and Backend/AI Engineer for this Smart India Hackathon 2025 project. We built a disaster response platform. I combined computer vision (OpenCV) and NLP pipelines for hazard classification, and mapped live disaster hotspots using DBSCAN spatial clustering with PostGIS.

OpenCV NLP DBSCAN PostGIS
Interactive Architecture

The Engineering Lab

SwiftPredict v2 is a pip-installable AutoML library that automates feature preprocessing, model search, training, and evaluation. Hover or click any node in the architectural flowchart below to inspect its script files and data flow.

User Interfaces
React/Next.js
SwiftPredictUI
Streamlit Dashboard
Testing Interface
Python SDK
SwiftPredict Client
CLI Interface
swiftpredict command
Backend Services
FastAPI Server
logger_apis.py
SwiftPredict Core
swift_predict.py
AutoML Engine
automl_trainer.py
Preprocessing
preprocessing.py
Data Storage
MongoDB
SwiftPredict Database
CSV Data Files
train.csv, test.csv
Architecture Details

SwiftPredict v2 Architecture

A pipeline that coordinates automatic data preprocessing, scaling, encoding, and model cross-validation. Client interfaces communicate with the core library while FastAPI manages asynchronous metric logging to MongoDB.

Hover or click any node on the left to inspect file-level details.

Keiro is a self-hostable adaptive RAG gateway proxy. It coordinates rapid local cache matching via a Golang router and routes cache-misses to Python classification and semantic retrieval gRPC services.

Natural Language Space
User Query
Prompt Entry
Go Gateway (Service: gateway)
Go chi Router
chi.Router (v1/query)
Ingestion Queue
IngestionQueue (Go Channel)
Semantic Cache
Semantic Cache (LRU)
Python Intelligence (Service: intelligence) & Storage
gRPC Classifier
ClassifyQueryType
gRPC Retrieval
ExecuteRetrieval
gRPC Generator
GenerateResponse
ChromaDB
Collections Storage
Result Space
Generated Answer
Natural Language Response
Architecture Details

Keiro RAG Gateway Architecture

A dual-language service orchestration. A high-concurrency Go proxy routing client requests through an LRU semantic cache, communicating via gRPC calls to a Python backend handling model-based routing and response synthesis.

Hover or click any node on the left to inspect file-level details.

Samudra Sachet is a real-time disaster monitoring and hazard alert system. Live media reports, Tweets, and citizen reports flow through a FastAPI gateway, store media on Cloudinary, run CNN and NLP validation pipelines, and perform DBSCAN hotspot clustering written to Firestore.

Frontend Layer - /frontend/
LandingPageProtoMultiLingual/
Multi-language public interface
Anonymous report submission
login-authorized/
Citizen authentication portal
Firebase Auth integration
PROTOWITHDASHBOARD/
Official command dashboard
Live clustering visualization
Backend Layer - /backend/
main/
FastAPI application
Route handlers + orchestration
Firebase Services
Authentication
Cloud Firestore
Storage
Cloudinary Integration
Media upload/retrieval
Image/video storage
Data Layer
MongoDB
Motor async driver
Validated schemas
Cloud Firestore
Real-time updates
JSON validation
External Services
Twitter API v2
point_radius geo-filters
Social media stream
ML Pipeline - /backend/ML_models/
data/
- Augmented_data.csv
- Embedded_dataset.csv
- enriched_urgency_seed_data.csv
PanicMeterModel
PyTorch CNN
Image urgency scoring
NLP Pipeline
GloVe embeddings
Ensemble regressors
Hotspot Detection
DBSCAN/HDBSCAN
cluster_data() function
Architecture Details

Samudra Sachet Architecture

A real-time disaster reporting and alert dispatcher. Anonymous and authenticated citizens stream crisis reports to FastAPI, validating text urgency via GloVe regressors, scoring image panic metrics with PyTorch CNNs, and clustering incidents using DBSCAN to Firestore.

Hover or click any node on the left to inspect file-level details.

Credentials

Certifications & learning

IBM AI Agents using RAG and LangChain

IBM
Mar 2025

Deep Learning with PyTorch

IBM
Feb – May 2025

CNN Specialization

DeepLearning.AI
2025
Contact

Let's build something

I'm actively looking for ML Engineering internships, AI backend roles, and research positions in RL and GenAI. If you're building something interesting, I'd love to chat.

mranjanjena253@gmail.com github.com/ManasRanjanJena253 linkedin.com/in/manasranjanjena253 +91 78887 22346
Thank you! Message sent. I will get back to you within 24 hours.
Terminal CLI `
manas@pec-chandigarh: ~ (type 'help' for commands)
Welcome to Manas Ranjan Jena's interactive portfolio CLI!
Type help to view all available commands.
 
visitor@portfolio:~$