// backend · cloud · ml infra
Sagar
Jaswal_
Software Development Engineer — Backend & Cloud
Two years of architecting and deploying scalable backend systems, GPU-accelerated video processing pipelines and ML-powered product features on cloud-native infrastructure. Passionate about ML infrastructure and open-source & enterprise cloud-native tooling.
01 work_experience.log
Experience
Reverse-chronological. Each entry is a deployment.
Clipo AI
Backend Engineer
- Architected and owned the backend platform end to end as the sole backend engineer, scaling it from pre-launch to 10,000+ registered users and 6,000+ exported videos.
- Developed the video editing and rendering engine directly on CUDA-accelerated FFmpeg with FastAPI, delivering 40+ customizable video templates, animated captions, transitions and real-time subtitle burning, with exported videos matching the frontend preview pixel for pixel.
- Engineered Auto-reframe to convert landscape video to vertical, tracking multiple faces across frames with the Hungarian algorithm, detecting the active speaker from mouth and jaw movement without audio, and applying multi-pass smoothing for stable framing.
- Implemented visual search using vision-language models to locate specific moments within videos, along with object-based reframing powered by Meta’s SAM 3.
- Deployed GPU workloads on Google Kubernetes Engine (GKE) with shared GPUs and per-workload memory limits, autoscaling between 1 and 5 GPU nodes, and Celery/Redis job queues, processing 10,000+ minutes of video every month.
- Automated observability across all pods with Prometheus, Grafana and Loki, and integrated PostHog product analytics to track both system performance and user behaviour.
- Profiled latency percentiles (p50/p95/p99) and resource usage through Grafana dashboards to identify blocking operations on critical paths, cutting response times from minutes to seconds for some workloads through parallelization, and from 14 seconds to under a second through targeted optimizations.
- Designed a payment system that handles international subscriptions and one-time payments seamlessly for users, with safeguards against duplicate charges, and introduced database caching that substantially reduced database load.
Panoculon Labs
Software Engineer
- Took full ownership of Panocle, the companion mobile app for the company’s smart glasses for visually impaired users and a core part of the product.
- Co-designed the Bluetooth Low Energy transfer protocol with the embedded engineers, splitting each image and audio capture into chunks with size headers, and implemented an app-side state machine that advanced only once a capture had fully arrived.
- Developed a context-aware AI assistant using Gemini with Retrieval-Augmented Generation (RAG), integrating function calling, vector search and contextual memory retrieval, along with a live mode for real-time conversation.
- Quantized YOLOv8 and BERT models so that currency recognition and query routing run directly on the phone.
- Launched a cross-platform meeting app in Flutter with real-time transcription, speaker diarization (Whisper), onboarding flows, tiered subscriptions, a referral system and payments via Apple IAP and Razorpay, scaling to 300+ users within 2 weeks.
- Architected its serverless backend on AWS (Cognito, API Gateway, Lambda, DynamoDB, Amplify) with asynchronous S3–Lambda workflows and WebSocket notifications.
- Migrated the user base to Firebase, halving latency, and reduced monthly AWS costs by 70% by eliminating stale resources and optimizing usage patterns.
- Created the company’s marketing site with Next.js (React + TypeScript), implementing SEO strategies and CI/CD workflows that improved load times and increased web traffic.
Panoculon Labs
Full Stack Developer — Intern
- Trained a YOLOv8 model for real-time Indian currency detection on a self-collected and labelled dataset spanning all seven denominations in old and new designs, applying data augmentation, transfer learning and hyperparameter tuning to achieve high precision and recall, even on folded and poorly lit notes.
- Fine-tuned a BERT model to route users’ spoken queries to the appropriate feature, implementing data processing, tokenization and model optimization to improve accuracy and reduce training time.
02 projects.d
Selected Builds
Personal & open-source. Code over slides.
Point of Sale App
↗offline-first · in live use
An offline-first billing app in Flutter with an MVVM architecture for a dairy and restaurant business, now in live use — digitizing all of the shop’s sales and billing records, with Bluetooth printing of kitchen order tickets (KOT) and Google Drive cloud backup.
Online Coding IDE
↗final-year project · team lead (4)
Led a team of four building a full-stack online IDE with Next.js and Tailwind CSS, featuring Docker-based sandboxed code execution for Python, JavaScript & C++ with container isolation, resource limits and real-time output, secure authentication and MongoDB-backed file management.
03 stack.toml
Stack
What I reach for.
languages
C++PythonDartJavaScriptTypeScriptHTMLCSS
cloud & devops
GCP · GKECompute EngineCloud StorageAWSLambdaAPI GatewayDynamoDBCognitoS3AmplifyKubernetesDockerCeleryRedisPrometheusGrafanaLokiFirebase
ml & media
YOLOv8BERTRAGVector SearchModel QuantizationVision-Language ModelsFFmpeg (CUDA)
frameworks & data
FastAPINode.jsExpress.jsReactJSNext.jsFlutterMongoDBPostgreSQLMySQLGit
04 signals.log
Leadership & Awards
001
5+ industry exhibitions — represented Clipo AI, presenting the product and building industry connections — product
002
Coding sessions for junior students through the college coding club — mentoring
003
500+ DSA problems solved across competitive programming platforms — algorithms
004
Hacktoberfest 2021 — open-source contributions to Flutter projects — open-source
005
2nd prize — Youth Parliament 2018 , Kendriya Vidyalaya No. 2, Kozhikode — public speaking
05 writing.log
Writing
Notes from shipping in prod.
06 contact.sh
Get in touch
Open to build. Drop a message and I'll reply.