Rishik Kasula
ML & data engineering. Shipped and measured.
ML and data systems, measured honestly — in dollars, against SEC filings, by what they prevent. Every project number on this site traces to a public repo.
ML and data systems, measured honestly — in dollars, against SEC filings, by what they prevent. Every project number on this site traces to a public repo.
fraudscore, filinglens, tickflow — public repos that grade themselves against verifiable ground truth and publish the intervals.
~3,500 lines of verified code on RefExp — structural explainability for code refactorings, built on Joern and RefactoringMiner.1
Data Science concentration at UNC Charlotte. 3.80 GPA. Eight A's across ten graduate courses.
Kafka + ZooKeeper real-time ingestion and Power BI dashboards for banking IT at Inspirisys Solutions.
8 certifications — AWS Data Engineering, DeepLearning.AI mathematics, IBM Python, Docker. All independently verifiable.
Two graduate appointments at UNC Charlotte and an industry internship in Chennai — each one shipping working systems, from a research pipeline over code property graphs to real-time data ingestion in banking IT.
Existing tools detect that a refactoring happened. RefExp answers the structural question none of them do: which other changes in the same commit are connected to it — via call paths, shared types, or data flow — built over RefactoringMiner and Joern Code Property Graphs. Joint research with a second RA — the contributions below are git-verified as mine.
Supported a graduate course spanning ML/DL model development, architecture for AI systems, data quality and provenance, explainability, and testing for AI. Weekly office hours; guided teams through a six-deliverable, GitHub-based research project — proposal to data collection to training, evaluation, and final report — and assisted with grading across the weekly paper-reading series.
Implemented an Apache Kafka + ZooKeeper pipeline architecture for real-time ingestion in a banking-IT context. Built Power BI dashboards tracking banking-IT KPIs. Deployed PrivateGPT in two modes — fully local LLM and OpenAI-API-backed — with Hugging Face, CUDA, and Poetry on Windows 11.
The three shipped projects share one thesis: a system should evaluate itself against verifiable ground truth — and publish what that evaluation actually shows, intervals and failures included. fraudscore prices fraud decisions in dollars with calibrated probabilities. filinglens grades LLM extraction against companies' own XBRL filings. tickflow is a streaming quality gate measured by the corrupted bars it prevents. Each has a full case study below; earlier academic work follows in brief.
A calibrated card-fraud scoring service that prices every decision: a transaction is reviewed when fraud probability × amount exceeds the cost of reviewing it. Served over FastAPI, evaluated on a chronological hold-out in dollars with bootstrap confidence intervals. One honest finding worth reading: a plain logistic model beat gradient boosting under honest chronological evaluation.
An evaluation harness measuring whether small local LLMs can be trusted to pull financial figures out of SEC 10-Ks. The trick: every US filer publishes its numbers twice — prose and structured XBRL — so the filing grades the extraction automatically, no human answer key. A 600-call deterministic grid (10 filers × 5 KPIs × 3 local models × 2×2 strategy ablation), every failure hand-labeled. It ships a measurement, not an extractor.
A contract-enforcement layer for Kafka-compatible streams: declarative per-topic contracts (Avro + six semantic rules in YAML) enforced inline, violations quarantined with self-describing envelopes, a downstream bar builder protected by a measurable SLO. Every quality claim is graded by seeded fault-injection replay against a ground-truth manifest in CI — deliberately uncatchable faults counted as misses, not excluded. All results are fixture-scale, and labeled as such.
The full story behind each shipped system — problem, approach, measured result, and the honest caveat. Click to expand.
Every skill below is backed by shipped code, graduate coursework, or a verified certification. Nothing padded. Nothing borrowed.
All certifications independently verifiable via Credly, Coursera, and Cognitive Class.
Open to ML and data engineering roles. Based in Charlotte, NC — happy to relocate or work remote.
1. Contribution volume verified against git commit history (author-attributed commits, RefExp v1 repository), corroborated by the co-RA's integration commit.
2. Internship credited as BIC4002 (10 credits) at VIT-AP University, graded S (Outstanding).
3. Measured on the held-out chronological test split of the ULB dataset; percentile bootstrap, B = 10,000, seeded. Full methodology and eval report in the fraudscore repository.
4. filinglens: headline-figure extraction with the financial statements already in the model's context; N = 10 large-cap us-gaap annual filers; cluster bootstrap by company, B = 10,000, seeded. Extraction with the document supplied is the easy end of the task — not comparable to any reasoning benchmark. Full methodology in the filinglens eval report.
5. tickflow: all figures are fixture-scale — produced from a committed, checksum-pinned synthetic fixture with a chosen ~2% fault mix, never live traffic. The project publishes no throughput or latency figure. Provenance for every number in the repository's committed telemetry artifact.