About
Hey! I'm Kalyan, and I love experimenting with agents across AI systems. I just finished my Master's in Computer Science from DePaul University, Chicago this June. I take pride in my expertise across the fields Computer Vision and LLMs.
At sensen.ai my computer vision models turned traffic footage into enforcement tickets that reached real people, so a digit misread in bad light became someone's actual problem. That taught me to distrust clean metrics and keep asking whether a system still holds 6 months after it ships, and whether I can prove it.
My grad research comes from the same instinct. The Multi-Agent Inference Reliability Framework is 2 linked studies: one asks whether a runtime monitor around an LLM actually helps, and finds the point where it starts to hurt; the other asks whether the automated judge inside an LLM pipeline can be fooled. Under review.
Away from the models I build the plumbing too: a streaming lakehouse with its own data-contract engine and an MLOps pipeline that fails a bad model in CI.
Experience
Graduate Research Engineer · DePaul University
Led the multi-agent inference reliability research with 2 faculty co-authors, now under review. Owned it from the problem to the architecture to the statistics, across 10 phases and 5 model families.
Read the research →Data Scientist · sensen.ai
Sole data scientist and forward-deployed technical contact across sensen's client base. Fine-tuned the ANPR vehicle-entry classifier (CNN, transfer learning) to ~90% accuracy, built the company-standard model-evaluation framework whose discrepancy metric tied model error to lost revenue, and owned 300+ pipelines plus the analytical SQL serving all 80 clients on AWS, Azure, and GCP.
Data Engineer · Aecom & Siri
Built Python and pandas pipelines for cost consolidation, cutting dashboard turnaround from 3 hours to under 1 and tuning SQL across cost databases from about 20s to 8s.
Projects
secret-agent
Curious what runs under the hood of Claude Code. Built the agent loop from scratch, no framework, 53 lines. A hostile review found a full sandbox escape, since closed. Step through a live trace.
GroundTruth
Telling a real effect from luck on a 14M-row A/B log. CUPED cut the noise and tightened the intervals, turning a borderline result into a clear signal. Try the simulator.
PredictOps
A model that looks great in a notebook isn't enough. A CI gate blocks a worse model, Evidently catches drift, and MLflow promotes on merit.
StreamLake
Bad incoming data poisons everything downstream. Kafka to Iceberg with a data-contract engine that quarantines any record breaking the rules. Break one and watch it get caught.