Gagandeep Shivanna

Remote

• Curate and author high-quality training data for LLM fine-tuning, writing gold-standard reference responses scored against a rubric spanning accuracy, relevance, and safety.
• Produce labelled preference data through structured side-by-side model comparisons, generating the ranked response pairs consumed downstream for reward-model training.
• Define annotation guidelines and intent-level task decomposition specs for multi-step agent workflows, raising inter-annotator agreement and dataset consistency across the pipeline.

• RLHF
• Model Evaluation
• Preference Data
• Agent Workflows
• Rubric Design

Berlin, Germany

• Built end-to-end RAG pipelines for document classification and information extraction, using ChromaDB for vector storage and semantic retrieval over domain document corpora.
• Tuned chunking strategy, embedding selection, and retrieval depth against a held-out evaluation set, improving extraction accuracy over a zero-shot baseline.
• Applied prompt engineering and structured output constraints to make extraction results machine-parseable and reproducible across model versions.

• RAG
• ChromaDB
• Semantic Search
• Prompt Engineering
• NLP

Bangalore, India

• Developed and maintained enterprise Java/Spring Boot applications (Zaxby's, Tonic): scalable RESTful APIs and backend microservices with database integration across MySQL and MongoDB.
• Architected modular backend services with Dependency Injection and AOP patterns to improve maintainability and reduce coupling across service layers.
• Led API debugging and unit/integration testing with JUnit and Mockito, identifying and resolving 50+ critical bugs and enforcing code quality through structured peer reviews.
• Automated deployment pipelines with Docker and Jenkins CI/CD, reducing manual release overhead and improving deployment reliability.
• Contributed to 15+ Agile sprint cycles, consistently meeting milestones and recognized for technical ownership and cross-functional collaboration.

• Java
• Spring Boot
• REST APIs
• Microservices
• Docker
• Jenkins

Bangalore, India

• Trained on Java and Spring Boot fundamentals, then contributed to enterprise client applications under senior engineer mentorship.
• Worked across multiple teams on shared modules, picking up code review etiquette, Git workflows, and Agile ceremonies.
• Converted to a full-time Backend Engineer offer at the end of the internship.

• Java
• Spring Boot
• Git
• Agile

SRH University of Berlin · Berlin, Germany

• Master's thesis: a failure taxonomy of LLM-generated Infrastructure-as-Code, a four-referee verifier (Terraform validate/plan, Checkov, tfsec, OPA/Rego) plus an automated self-repair loop across 560 runs and three models, with survival analysis over 2,068 logged errors.
• Coursework across machine learning, deep learning, cloud, and data engineering; hands-on with AWS, Azure, and GCP, backed by 30+ cloud and AI badges.
• Current CGPA 1.4 on the German 1.0–5.0 scale (1.0 = best).

BMS College of Engineering, Bangalore · Bangalore, India

• Final grade 8.17 / 10 (EQF Level 6).
• Projects: NLP language detection (Naive Bayes), biomedical waste classification (ANN/VGG16), and a full-stack hospital management system (PHP/MySQL).

Non-technical users can't query data without SQL.

Production-deployed analytics platform with a natural-language-to-SQL pipeline, sandboxed query execution, real-time SSE streaming, and auto-generated Recharts visualizations. Includes a Gaussian Copula synthetic-data engine with fidelity metrics and a text-to-SQL eval harness with execution-accuracy benchmarking tracked in MLflow.

Key decisions: Two-model Claude Sonnet/Haiku split for cost/latency; DuckDB for in-process analytical queries; execution-accuracy eval harness over exact-match.

datagrid.gagandeep.app · github.com/GagandeepS017/Datagrid

• FastAPI
• React
• DuckDB
• Claude Sonnet/Haiku API
• MLflow
• SDV
• Recharts
• Vercel
• Render

Live meetings are expensive and exclude async/distributed teams.

A SaaS platform that replaces live meetings. Teams submit input asynchronously through a guided chat powered by Claude Sonnet, which asks follow-up questions on its own, then synthesizes all responses into a structured decision summary on a real-time WebSocket dashboard.

Key decisions: WebSocket dashboard for live state; Claude-driven adaptive follow-up questioning.

• FastAPI
• React/TypeScript
• Claude Sonnet API
• WebSockets
• Python

Manual AWS account setup is slow, inconsistent, and error-prone.

An Infrastructure-as-Code pipeline that provisions complete AWS accounts from a single YAML config: IAM users, groups and policies, S3 buckets, and VPC networking, all through Terraform and GitHub Actions with OIDC auth. It also writes a per-user runbook after every deploy.

Key decisions: OIDC over long-lived keys; single-YAML config as the source of truth.

github.com/GagandeepS017/aws-onboarding-automation

• Terraform
• GitHub Actions (OIDC)
• AWS
• Python

Self-repair improves LLM-generated Terraform, but nobody has mapped which failures it fixes and which survive every retry.

An empirical study of whether LLMs can self-repair the Terraform they write. A pipeline generates Terraform, checks it against four independent referees (syntax/plan, security via Checkov + tfsec, org policy via OPA/Conftest, and a completeness check), then feeds the errors back for up to 10 repair rounds: 560 runs across 50 scenarios, 3 seeds, and three models, logging 2,068 errors down to the rule and round. The headline finding is a capability ladder, not a single aggregate number: Claude Sonnet 5 resolves 95% of tasks and fails only on unfamiliar resource types, GPT-4o resolves 53% and fails specifically on cross-resource reasoning like networking, and Llama 3.1 resolves 9% and fails to converge at all. A controlled experiment isolating feedback quality as the only variable shows detailed error messages causally drive repair success (McNemar p < 0.0001), and a repair-integrity classifier that diffs code before and after each fix found the frontier models never cheat by deleting a required resource, only the weakest model does. Reproduced against the IaC-Eval (NeurIPS 2024) benchmark: the one-shot GPT-4o score (21%) matches their published 19.36%, and repair lifts it to 45%.

Key decisions: Prompt-based repair, no fine-tuning, to isolate feedback as the sole variable; four independent referees, including a completeness check, so a fix can't be faked by deleting the flagged resource; a 10-iteration cap matching prior work (IaCGen) for comparability; two benchmark tiers, ported IaC-Eval tasks plus reverse-engineered production scenarios, with calibrated gold references.

• Python
• Terraform
• Checkov
• tfsec
• OPA/Conftest
• GPT-4o / Claude Sonnet / Llama 3.1
• React
• Statistics (McNemar, survival analysis)

GitHub MCP Server(Private)

Remote MCP server on Cloudflare Workers that lets Claude on my phone browse my repos and file ideas as issues. OAuth 2.1 with PKCE, a server-side PAT (no token passthrough), and a deliberately read-heavy toolset: nine tools, one write.

TypeScript · MCP · Cloudflare Workers · OAuth 2.1 · Octokit

ChessTalk(Private Beta)

Chess improvement app that pulls your chess.com games, runs Stockfish in the browser to classify blunders by game phase, and coaches you through your weak spots with talking pieces.

React · Vite · Stockfish WASM · chess.com API · Cloudflare Workers

Deepfake Detection in Art & Media

CNN- and GAN-based models that flag AI-manipulated images, separating authentic artwork from synthetic forgeries.

Python · TensorFlow · CNN · GANs · OpenCV

YouTube AI Video Discovery

Semantic enrichment pipeline that surfaces under-described or less-popular videos using NLP and Google Cloud AI APIs.

Python · NLP · GCP · Vertex AI

Language Detection (NLP)

Text-based language identification for Indian languages using a Naive Bayes classifier over n-gram features.

Python · Scikit-learn · NLP · Naive Bayes

Biomedical Waste Classification

COVID-era biomedical waste segregation using an ANN with VGG16 transfer learning for image classification.

Python · Keras · VGG16 · ANN