Harmanpreet Singh
Full-stack software engineer and applied ML researcher bridging high-scale distributed systems, real-time web architectures, and ambient sensor intelligence.

[ Background & Story ]
How I got here
I am a software engineer and researcher focused on the intersection of full-stack systems, distributed architectures, and applied machine learning. I believe the best software balances rigorous evaluation, clear abstractions, and end-to-end craftsmanship — whether that means shaving 400 milliseconds off a high-volume database query or training specialized AI models that run with zero inference bloat.
My foundation started in Computer Engineering at Savitribai Phule Pune University, diving deep into operating systems, compilers, networks, and algorithms. My undergraduate thesis was a resume ranking engine powered by NLP and ML; I led the team, built the full-stack architecture, and published our findings in IJSREM.
In industry, I engineered systems for high-stakes, production environments. At Lykan Technologies, I optimized asynchronous FastAPI backends, designed background workers handling 10,000+ records daily, and tuned Kubernetes deployments on AWS — reducing endpoint response times by 35% and infrastructure costs by 15%. At Stylabs AI Native Studio, I built applications serving over 500,000 users across commercial marketplaces like Manzil and Hireavilla, architected real-time WebSocket chat engines across multiple stakeholder roles, and built 40+ MongoDB aggregation pipelines that cut query latency by 53%.
To deepen my theoretical and systems work, I joined the University of Georgia for my Master of Science in Computer Science, advised by Dr. Fei Dou. My research tackles ambient sensor intelligence and multimodal self-supervised learning on CASAS smart homes — recognizing daily human activities without invasive cameras or microphones. By factorizing sparse sensor windows into sequence, narrative text, floorplan raster, and graph views with local competence-weighted fusion, our work achieved the highest accuracy across all 5 testbed homes and beat the strongest multimodal baselines at one-sixth of the computational cost (submitted to IEEE PerCom 2027).
I don't leave research in Jupyter notebooks. To bring my thesis model into the real world, I engineered HomeSense — an end-to-end digital-twin platform streaming live sensor telemetry through Apache Kafka, orchestrating automated anomaly detection with human review, embedding a LangGraph assistant, and deploying everything onto Kubernetes with Helm. Alongside that, I build developer-first educational tools like SystemForge (a distributed systems simulation lab used by 10,000+ engineers) and MLForge (an in-browser ML laboratory implementing neural networks from scratch in TypeScript).
[ Philosophy & Standards ]
How I work
Evaluation is part of design.
Every model and system is measured against standardized stratified splits, real-world workloads, and multiple random seeds rather than convenient headline numbers.
Inspect the raw data.
Before trusting an aggregate score, I build playback tools and visualizers to inspect raw event streams, coordinate rasters, and boundary cases directly.
Measure before optimizing.
Profile real production workloads and query plans before making architectural bets — like dropping MongoDB aggregation latency from 750 ms to 350 ms.
Write clear contracts.
Self-documenting APIs, strict TypeScript schemas, and thorough documentation ensure frontend, backend, and mobile developers can ship fast in parallel.
End-to-end ownership.
From Kafka streaming pipelines and database schemas to reactive interfaces, digital twins, and Kubernetes Helm deployments, I own the entire lifecycle.
Build with observability.
Instrument systems with structured logging, health checks, anomaly review queues, and latency metrics from commit zero.
[ Publications & Surveys ]
Research & writing
Multimodal Self-Supervised Activity Recognition on CASAS Smart-Home Sensor Streams
M.S. thesis, University of Georgia
Four derived views of sensor windows, evaluated on four CASAS houses.
FACET: Factorized Ambient-Sensor Experts with Local Competence-Weighted Fusion for Smart-Home Activity Recognition
Conference paper, submitted and under review (joint work)
Not yet accepted or published.
Survey on Contrastive Self-Supervised Learning in Human Activity Recognition
Graduate research seminar, University of Georgia
Sole-authored, ACM-style survey of contrastive approaches (SimCLR, CPC, multi-sensor fusion, time-frequency contrast) and open problems such as false-negative handling and augmentation design.
Resume Ranking System Using Machine Learning & NLP
International Journal of Scientific Research in Engineering and Management (IJSREM), Vol. 07, Issue 05, May 2023
Written with the undergraduate project team.
I also volunteered at the University of Georgia School of Computing 2025 Research Day, where I compiled the book of abstracts for the research seminars and academic posters on display.
[ Get in Touch ]
Elsewhere
Based in Atlanta, Georgia. Open to conversations about software engineering, distributed systems, and applied machine learning roles.