Under KV-cache compression, the model's behaviour degraded before the safety probe did.
Qwen3KV-cache compressionLinear probesSnapKV
Rahul Krishnan
MSc Data Science student and researcher at Universitaet Trier. I build efficient and reliable machine learning systems, currently focusing on memory-efficient LLM inference.
Ask about my work
A retrieval assistant indexed over my papers, projects, and CV answers questions here.
Retrieving context...
Generating answer...
I am a machine learning researcher and software engineer specializing in efficient LLM inference, numerical optimization, and reliable AI systems. At Universitaet Trier, I research KV cache compression under Prof. Dr. Volker Schulz, pairing graduate research with more than four years of enterprise software development experience at SAP Labs India.
My primary research tackles the memory wall in long-context language models. Through JoLT, an approach combining partial Tucker decomposition with low-bit residuals, I demonstrated near-lossless 2x to 3x KV cache compression across five models—maintaining accuracy within 2 percentage points on 64K-context and GSM8K benchmarks, with a custom fused kernel reducing peak decode memory by 3.61x. I also study model trust, exploring whether step-by-step explanations reflect actual model reasoning and analyzing how asymmetric penalties in RLVR steer model calibration.
Having designed large-scale distributed systems and compliance workflows at SAP, I hold machine learning to high engineering standards. I build complete systems—including a production RAG assistant on Google Cloud Run and a sub-100ms Two-Tower recommender on Azure—focused on rigorous empirical validation, measured performance, and transparent limitations.
Experience
2026
Research Assistant / Student Assistant (studentische Hilfskraft)Universitaet Trier, Department of Mathematics. Trier, Germany
- I researched KV cache compression under Prof. Dr. Volker Schulz, co-authoring the JoLT preprint that reports near-lossless 2x to 3x compression across five models and four model families: https://arxiv.org/abs/2607.12550.
- I graded weekly Numerical Optimization exercise sheets and provided technical feedback on unconstrained and constrained optimization, line search, trust-region methods, KKT conditions, and numerical linear algebra.
Mar 2024 - Sep 2024
Software Developer (SDE II)SAP Labs India Pvt. Ltd. Bengaluru, India
- I designed a rule-based, declarative content-matching framework for tax-document field comparison over SAP iFlow, letting country teams configure field comparisons with no custom code. It reduced new-country onboarding from about 10 days to about 3 days and was adopted by the Spain, India, and Chile teams.
- I designed an iFlow bulk-retrieval algorithm for the DCC Spain reporting path, replacing about 100 paginated API calls with one bulk request and reducing report generation time by about 30%.
- I proposed a per-customer self-correcting ML system for content-matching results, with in-tenant training to avoid central data pooling. I presented the concept at an internal SAP innovation forum in early 2024; it was accepted as a future direction but was not shipped before my role ended.
- I was promoted from SDE I to SDE II after consistent top reviews and selected for SAP Labs India's fast-track architect career-development program, as the only person from my area in the program at the time.
Jul 2020 - Mar 2024
Associate Software Developer (SDE I)SAP Labs India Pvt. Ltd. Bengaluru, India
- I owned the result-processing pillar of the Document Consistency Cockpit end to end. I designed its schema, decoupling layer, factory-based extensibility, idempotency, and retry semantics. I load-tested it at 10M to 100M document cycles and deployed it for Spain and India.
- I initiated and led a Turkey error-handling overhaul for SAP DRC, categorizing two years of incidents by root cause and shipping fixes, clearer messages, and customer-facing FAQs. Incident volume fell about 88%, from roughly 50 to 2 or 3 per month, while the support rating rose from 2 or 3 to 4 or 5 out of 5.
- I re-engineered a failing daily Turkey vendor reconciliation job by replacing a timing-out database join with a hash-map lookup keyed on tax ID, eliminating recurring failures as the registered-entity list grew.
- I designed GDPR-aligned Data Protection and Privacy compliance for the SAP Document Consistency solution and delivered legal and feature updates across Turkey, Mexico, and Chile. I received the SAP Stellar Award in 2023 for the DCC and Turkey work.
Jan 2020 - Jul 2020
Software Engineering InternSAP Labs India Pvt. Ltd. Bengaluru, India
- I shipped backlog features for the SAP eDocument solution, now SAP DRC, including Turkey delivery-note support and an XML proxy update for upstream format changes.
- I delivered consistently ahead of deadline with near-zero QA defects and earned a direct full-time offer on the same team.
Education
Oct 2024 - Oct 2026
MSc Data ScienceUniversitaet Trier
2016 - 2020
BTech Computer Science and EngineeringAmrita Vishwa Vidyapeetham
A probe can flag when a model is likely to follow a misleading hint, but neither probes nor an LLM monitor reading the CoT showed CoT-specific evidence that the hint caused the answer.
Qwen3Linear probesActivation steeringBIG-Bench Hard
arXiv preprint, in preparation for Numerical Linear Algebra with Applications (NLAA), 2026
A companion to the JoLT paper that treats the KV cache as a fourth-order tensor, measures its singular-value spectra on Mistral-7B-v0.3 and LLaMA-2-13B, and compares Tucker, CP, tensor train and t-SVD at matched storage.
Result
Token and feature modes are low-rank while head and layer modes are nearly full-rank. Tucker has the lowest reconstruction error at every compression ratio from 2x to 5x, and post-RoPE keys lose 41% to 64% of their pre-RoPE compressibility.
arXiv preprint, 2026
My MSc thesis preprint reports near-lossless 2x to 3x KV cache compression, with accuracy staying within 2 percentage points of the uncompressed model, across five models and four model families.
Result
At 3x compression GSM8K moves -1.21 pp on Mistral-7B and RULER 64K moves -0.10 pp on LLaMA-3.1-8B, with perplexity within 0.3% of dense on all five models. The FlashJoLT fused decode kernel cuts peak decode KV memory by 3.61x.
Punishing wrong answers made the model skip more often, not more wisely.
GRPORLVRQwen 2.5LoRA
72.5% pairwise accuracy and 95% precision on 500 story-comparison test cases spanning five languages
PyTorch GeometricHeteroConvGPT-OSS 20BCORAL
Under severe temperature mismatch between model and reality, CVaR control cut constraint violations from 100% of scenarios to 0%
CasADiIPOPTCVXPYMPC
Top-3 finish in all three university Kaggle challenges
PolarsScikit-LearnPyTorchGemma-3-12B
RAG Resume API2025
Powers the live AI Q&A assistant embedded in this portfolio site
LlamaIndexFastAPIHybrid RetrievalOpenRouter
Two-Tower MLP reached NDCG@10 of 0.3342, up from 0.3051 for the baseline, with sub-100ms CPU inference
PyTorchSBERTFAISSAzure
F1 = 0.81 on 10,000+ imbalanced YouTube comments, CI/CD to Docker Hub
BERTPyTorchFlaskGitHub Actions
Reflective CoT reduced bias susceptibility by about 7 percentage points over standard CoT across 650 cases
Gemini 2.5 Flash LiteLLaMA 3.2 3BBIG-Bench Hard
International Conference on Communication and Signal Processing (ICCSP), 2020
Undergraduate research on automated glaucoma screening from retinal fundus images using optic-disc and optic-cup segmentation with SVM classification.
Result
Achieved an F1 score of 86% on the Drishti-GS dataset for SVM-based glaucoma detection.
Under KV-cache compression, the model's behaviour degraded before the safety probe did.
A probe can flag when a model is likely to follow a misleading hint, but neither probes nor an LLM monitor reading the CoT showed CoT-specific evidence that the hint caused the answer.
Near-lossless 2x to 3x KV cache compression with accuracy within 2 percentage points of the uncompressed model
Punishing wrong answers made the model skip more often, not more wisely.