Projects
From distributed message brokers to self-correcting AI pipelines. I design, build, and ship complete systems.
Open to opportunities
Manas Sharma
Software Engineer in my final year of Computer Science at Manipal. Recently shipped fintech backend infrastructure at RupeeFlo. I work where distributed systems meet production ML.
Tech Stack
Beyond the Code
Jpop Music Reading Fiction Otaku Gaming Sports
Experience
•
Backend Engineering Intern
RupeeFlo (Fintech)
Worked on NRI banking & investment infrastructure. Migrated document-generation (notarization, e-sign) pipelines from Supabase into the core Django backend, collapsing a 3-call chain into one API and cutting processing time by 80%, with 0% divergence proven pre-cutover using a PDF-diff parity harness. Designed a 3-tier OTP fallback (WhatsApp → SMS → voice) with Celery and Redis, using atomic SETNX locks, rate limits, and idempotent escalation, serving 15,000+ users and validated across 27 replay/race edge cases. Shipped an AI support draft generator to production: Freshdesk history and Customer-360 data fed into structured LLM tool-use calls, cutting reply drafting time ~90% with 70% coverage on 78 real tickets.
•
Machine Learning Intern
Reliance Industries Limited
Built real-time occupancy detection using YOLOv11 and DeepSort across 10+ concurrent RTSP streams, achieving 30+ FPS throughput with sub-second latency on GPU deployment. Designed a Multi-Camera Manager with spatial deduplication and priority-based conflict resolution to eliminate duplicate counts across overlapping FOVs. Implemented adaptive tracking with motion blur detection, improving accuracy by 25%, plus a FastAPI backend with WebSocket real-time updates and SQLAlchemy persistence.
Projects Completed
From distributed systems to AI/ML pipelines,
here are some of the production-grade systems I've built.
AI Chair Occupancy Analytics
Real-time chair occupancy detection using YOLOv11 and DeepSort with multi-camera spatial deduplication, hitting 30+ FPS across 10+ concurrent RTSP streams.
Driving Scene Description Generator
An automated VLM pipeline that generates structured scene descriptions from autonomous driving images. Compares 8 prompt strategies across 9 metrics on 100 BDD100K scenes (800 evaluations) with statistical significance testing, plus an AI agent for error analysis.
PriceScope
An end-to-end multimodal deep learning system for marketplace price prediction. Trained on 1.48M Mercari listings with a BiLSTM + Attention + MLP fusion model achieving 0.420 RMSLE, beating XGBoost/LightGBM baselines and matching a fine-tuned DistilBERT with 4.5x fewer parameters. Full stack with FastAPI, Next.js, and MongoDB.
Fleet Vehicle Data Gateway
A fully deployed real-time fleet monitoring platform with FastAPI, Next.js, WebSocket live updates, OAuth 2.0 authentication, and geofencing. Live demo included.
NeuralRAG
A self-correcting Retrieval-Augmented Generation system with LLaMA3, ChromaDB, and hybrid BM25+vector search. Features hallucination detection and automatic query refinement.
Distributed Message Broker
A high-performance, fault-tolerant distributed message broker in Go with HashiCorp Raft consensus, partition sharding, ISR replication, gRPC transport, and topic-based pub/sub. Measured 262K msg/s produce with fsync on and 490K msg/s consume.