Ajay
Kannan.
I build production data platforms, ML pipelines, and generative AI systems — where reliability matters as much as intelligence.
The short version
Computer Scientist working at the intersection of data systems and applied ML. Infrastructure-first, model-curious.
Interested in how LLMs behave when you put them behind a latency budget, an SLA, and a bill.
I'm an ML Systems and Backend Infrastructure Engineer currently at Pacific Gas & Electric, where I design and maintain the data platform that moves utility operations data at scale.
My work lives at the seam between classical data engineering and applied machine learning — scalable pipelines, observability, anomaly detection, and increasingly, LLM-driven tooling that actually runs in production.
I hold an M.S. in Computer Science from Arizona State University, with research in few-shot learning for TCR–Epitope affinity prediction. Before that: Microsoft, Sigtuple, and Solarillion Foundation — where I started with IoT and gesture recognition.
Where I've built
Pacific Gas & Electric
- Designed and deployed a production data platform (Electric Productivity Tracker) for large-scale utility datasets — improved throughput ~20% and reduced system errors ~15%.
- Engineered scalable, fault-tolerant data pipelines integrating GIS, Postgres, and internal services with consistent low-latency access.
- Built automated backend validation and testing infrastructure in Python, cutting deployment verification time by 50%.
- Applied ML and generative AI techniques in production to improve anomaly detection reliability, reducing false positives ~30%.
- Shipped observability and alerting systems (health checks, dashboards) that cut issue detection and resolution time by ~50%.
Arizona State University
- Developed contrastive loss for few-shot learning in TCR–Epitope affinity prediction, improving accuracy by 30%.
- Implemented a Siamese network to evaluate the efficiency of different K-values in few-shot learning.
Sigtuple Inc.
- Applied image processing to biological data using Siamese Networks, GANs, and Neural Style Transfer.
- Used alpha blending to recreate histology slide images — hitting 94% enhancement accuracy in image blending.
Microsoft
- Identified sensor-data patterns to improve crop yields (FarmBeats) using Python, ML, REST APIs, and PyTorch; built a heatmap generation system over GIS data.
- Led the Shopping on Cortana project — integrated Cortana with a shopping platform (Cassandra + C#, MS-IDC Foundry Team).
Solarillion Foundation
- Designed static and dynamic American Sign Language gesture recognition systems — 94–97% accuracy using accelerometers, Arduino, and ML (Extra Trees, Random Forest, Ridge Classifier).
- Led the IoT research group, mentoring junior researchers as a teaching assistant.
Selected projects
Tools of the trade
Languages
- Python
- Java
- C++
- C#
- JavaScript
- SQL
- Scala
- Node.js
ML / DL
- PyTorch
- TensorFlow
- Keras
- Hugging Face
- LangChain
- Scikit-learn
- NumPy
- Pandas
Backend
- Spring Boot
- Express.js
- REST APIs
- GraphQL
- Microservices
Data / Storage
- PostgreSQL
- MongoDB
- Cassandra
- DynamoDB
- Redis
- Spark
- Kafka
Cloud / DevOps
- AWS
- Azure
- Docker
- Kubernetes
- Git
Predictive modeling · Big Data · Decision Analytics · Exploratory Data Analysis · Arduino · HTML/CSS · Problem solving
Papers & posters
Few-Shot Learning for TCR–Epitope Binding Affinity Prediction
Let's
build
something.
Open to conversations about ML infrastructure, generative AI in production, or anything in the neighborhood.