ANIRUDHA KUCHIBHOTLA · AI & DATA ENGINEER
02 / EXPERIENCE
Three roles across enterprise backend, data pipelines, and production AI.
Building production AI, data, and analytics systems across retrieval, evaluation, performance analysis, backend services, distributed workflows, and document automation.
Worked across batch and streaming data pipelines, analytics, data quality, and cloud-based data systems.
Built and supported enterprise backend and data workflows across APIs, SQL, ETL, reporting, and production troubleshooting.
03 / SELECTED WORK
Three systems across AI engineering, data engineering, and analytics: how they were built and what changed.
Engineers frequently report the same underlying bug using completely different terminology, file paths, or error symptoms. Simple keyword search misses these connections because the vocabulary doesn't match.
Compared keyword-based (BM25), dense semantic (sentence embeddings), and hybrid reciprocal-rank fusion over historical GitHub issues to see which approach finds related problems more reliably.
Hybrid retrieval increased top-5 recall by +26.6% over pure keyword search by pairing exact keyword precision with conceptual semantic matching.
Evaluation context: 31 temporally valid duplicate pairs from Apache Airflow.
GitHub issues keep changing after they're filed: new comments, labels, duplicate markers, and fixes. If an evaluation uses that later information, results look better than they really are.
A pipeline that pulls issues and their history from the GitHub API, cleans them into a consistent format, links duplicates and fixes, and filters every dataset by date so evaluations never see the future.
9,059 from Apache Airflow and 6,413 from VS Code, processed into reproducible datasets.
Powers the retrieval results in the DevSignal AI project.
Transaction amounts alone tell an incomplete story. Understanding modern commerce requires linking consumer demographics, category baskets, payment rail choices, and return patterns into clear retail strategies.
An interactive exploration analyzing consumer transaction records across generations, product categories, and checkout channels, mapping multi-factor transaction behavior across five distinct customer stages.
High-intent digital wallet checkouts produced 35% faster purchase velocity, while installment options (BNPL) expanded basket sizes but carried distinctly higher return and fee sensitivity.
Analysis structure: 5 customer stages from category selection to payment incentive matrix.
Data collected from multiple sources, then cleaned, processed, and merged before analysis.
04 / How I work
Define the metric and measure the simple approach before building the clever one.
Retries, state, recovery, and observability belong in the architecture, not in the postmortem.
Define expected behavior, build repeatable tests, and make regressions visible before claiming improvement.
If a dashboard or report doesn't change what someone does next, it isn't finished.
05 / TECHNICAL CAPABILITIES
The tools behind the work above, grouped by where I use them. Not exhaustive; I pick up what the problem needs.
06 / WRITING
Short articles about technical decisions, experiments, and things I've learned while working through engineering problems.
Most AI features don't need a custom model. A practical order of decisions, from APIs to distillation.
Every call that ran returned 200 OK. The workflow still failed. Designing for the gaps between requests.
07 / ABOUT
Most of my work sits where data meets AI: getting data into shape, building systems on top of it, and measuring whether they actually work.
I also build open-source projects like DevSignal and write about the engineering decisions behind them.