resume
AI/ML and software engineer, studying for an M.Sc. in Data Science & AI at Saarland University. I build retrieval-augmented and agentic LLM systems end to end, from ingestion to evaluation to production, and I measure them. In production at DHC Business Solutions, and in the open with ProofRAG. Open to full-time AI/ML and software engineering roles from 2026/27.
experience
Werkstudent, Applied AI DevelopmentApr 2025 – Present
InnoLab · Saarbrücken
- Building knowledge graphs in Neo4j and generating answers from them, with retrieval strategies that keep every claim grounded and tied to a source citation.
- Designed and shipped agentic RAG workflows (LangChain, LangGraph, Azure GPT-4o) over heterogeneous enterprise documents, with hierarchy-aware chunking and auto-captioned tables and images indexed as separate references.
- Built an LLM-as-a-judge evaluation pipeline scoring answers on six criteria (groundedness, correctness, completeness, citation quality, overreach, refusal) alongside NDCG, recall and precision at k, used to decide between architectures and to benchmark models, including how reliable smaller models can be made.
- Compared GraphRAG with dense retrieval plus reranking and moved to the latter, cutting response latency 6–10× (55–90 s to 5–10 s).
- Exploring TypeSafe's Jev, a decision model trained with RLCD (reinforcement learning for calibrated decisions), to speed up existing LLM workflows with minimal cost to reliability.
more from this role
- Switched embeddings from ada-002 to text-embedding-3-small, 5× cheaper and better on German retrieval.
- Asynchronous scraping and entity extraction pipelines with bounded concurrency, processing 2,000+ pages across domains.
- Knowledge graphs from structured and unstructured sources using BAML and LangExtract, stored in SQLite where a graph database was unnecessary.
- A dual-index (knowledge base + solution archive) multilingual RAG system with ChromaDB and Jina Reranker.
- LangGraph pipelines for Jira ticket analysis: priority classification, root-cause extraction and workarounds, with Pydantic-structured outputs.
- Model selection by task fit (GPT-4o mini outperformed GPT-5 nano on our tasks), a versioned prompt library, Gradio demos for stakeholders, and OpenWebUI integration across local and hosted models, deployed with Docker.
Student Assistant, Full-StackApr – Jun 2025
Smart Service Engineering · Saarland Univ.
- Built the research group's website end to end in Django, with a role-based admin dashboard and scheduled, event-triggered backups; in production at iss.uni-saarland.de.
Software Analyst InternMar – Sep 2023
Government & Public Services · Mumbai
- Built REST APIs in Spring Boot and Node.js for government-sector applications and prototyped a React chatbot; contributed to an Oracle-to-PostgreSQL migration that sped up queries by ~25%.
Machine Learning Intern, Data PrivacySep 2021 – Feb 2022
PRISM program · Bengaluru (remote)
- Built a PII detection framework with 18 custom NLTK and spaCy recognisers at 92% detection accuracy, with masking, tokenisation and redaction, served as a Django REST API. The team received Samsung R&D's Certificate of Excellence.
projects
- Tether — Creator & Android Engineer, Tendril Works (May 2026 – present). A local-first Android reader for articles, feeds, PDFs and EPUBs, built in Kotlin and Jetpack Compose with a signed release pipeline; released on Google Play with no accounts, ads or analytics.
- SamplerScope — Top 10% at Apart Research's Digital Minds sprint (Aug 2026). A toolkit that isolates how decoding rules change language-agent behaviour with logits held fixed, computing outcomes exactly across 11 decoders; greedy decoding raised benchmark return in 12 of 24 conditions and lowered it in the other 12.
- ProofRAG — Open-source Python CLI and library on PyPI, with a GitHub Action and agent skill: generates golden sets, scores retrieval and answers, and gates regressions in CI. Swappable LLM providers and scoring backends, strict validation of judge output, run fingerprints that refuse incompatible baseline comparisons, 188 tests on Python 3.11–3.14 and trusted publishing to PyPI.
- Program Repair Hints — Raised Phi-3-mini's repair pass rate from 36% to 88% with LoRA fine-tuning, and evaluated hint quality for correctness, informativeness and concealment.
- Fact Check LIAR — Hybrid BM25 and FAISS retrieval with a fine-tuned six-class BERT classifier, cached indexes and a structured fallback when no local LLM is available.
Also: rag-utils, prescriptive bias in LLM sampling, data selection for PEFT, scribble-supervised segmentation, healthcare utilization, Hukka Bukka and VoxArt.
open source
- zed-industries/zed (merged, Rust) — JSONL and NDJSON support in the tabular preview, with source-line errors, recovery after edits, regression tests and a CSV row-deletion crash fix. Write-up.
- neo4j-labs/create-context-graph — #33 (merged): connection-lifecycle fix with tests · #51 (open):
--ontology-fileCLI option with validation and tests. - microsoft/essex-toolkit (open) — Strip markdown code fences before JSON parsing, with tests.
education
M.Sc. Data Science & AIOct 2023 – Present
Saarbrücken, Germany
- Focus on NLP, neural networks, generative AI and machine translation; seminars on LLMs and on compositionality in neural networks.
B.Tech. Computer ScienceAug 2019 – Jun 2023
Mumbai, India
- Silver Medal (2nd in graduating class) and the Shri Baljit Shastri Award for academic excellence and values.
- Thesis: voice-to-image generation with NLP and sentiment analysis.
skills
- AI & ML
- RAG and agentic workflows (LangChain, LangGraph), LLM evaluation (LLM-as-judge, Ragas, DeepEval), PyTorch, Hugging Face Transformers, PEFT/LoRA, FAISS, BM25, rerankers, knowledge graphs
- Languages
- Python, TypeScript, Java, Kotlin, SQL
- Backend & data
- FastAPI, Django, Spring Boot, PostgreSQL, Neo4j, ChromaDB, SQLite
- Web & mobile
- Angular, React, Next.js, Android with Jetpack Compose
- Tools
- Git, Docker, GitHub Actions, Cloudflare Workers, Linux
- Spoken
- English (C1), German (A2, learning), Hindi (native)