EECS @ UC Berkeley
Yash Thapliyal
My work centers around building AI systems that are trustworthy, efficient, and useful.
- EECS @ UC Berkeley
- Currently @ Berkeley AI Research (BAIR)
- Prev @ EleutherAI, Sky Computing Lab, Splunk AI, Google, Amazon, C.Light Technologies, and smartQED
- Interested in Agentic AI, AI Safety & Alignment, Multi-Agent Systems, Reinforcement Learning, AI Memory, and Post Training
Work
Where I've worked and what I've built.
EleutherAI
Jul 2026 - Sep 2026
AI Safety Research Fellow, SOAR-S3
Remote
- Designed an evaluation harness for deceptive compliance in tool-using agents, replay-verifying tool calls against fresh environments to detect false completion claims across 481 trajectories, 8 models, and 15 pressure conditions.
Splunk (Cisco)
May 2026 - Aug 2026
Machine Learning Engineer Intern, Splunk AI
San Jose, CA
- Architected a multi-tier memory system for AI agents spanning short-term, session, and long-term memory, extracting durable facts, user preferences, decisions, and workflow patterns from conversations, documents, and tool traces.
- Distilled a 120B teacher into a 3B on-device student (40x fewer parameters) with MLX/LoRA; explored preference optimization and RL post-training via DPO/GRPO, alongside quantization and data scaling, to identify the best-performing training recipe; built a leakage-resistant evaluation framework with MiniLM embeddings and bipartite matching.
- Built a prompt-optimization platform from 0 to 1 (Docker, FastAPI, SQLAlchemy, dashboard, CLI, MCP) using GEPA via DSPy, lifting task accuracy by up to +50 points (0.27 to 0.77) across 8 optimization jobs.
Sky Computing Lab
Apr 2026 - Sep 2026
AI Researcher
Berkeley, CA
- Built a FAISS-based reasoning-trace retrieval pipeline over 58K chain-of-thought traces, evaluating whether retrieved solutions from related problems improve LLM mathematical reasoning across 7,680 graded generations on AIME 2025–26.
Berkeley AI Research Lab (BAIR)
Oct 2025 - Present
AI Researcher
Berkeley, CA
- Built an end-to-end synthetic speech pipeline generating 1,700+ validated minimal-pair clips with phoneme-level supervision across three voices using eSpeak and Piper TTS.
- Fine-tuned Qwen2.5-Omni-3B with LoRA/PEFT to transcribe speech as spoken rather than “correcting” unusual words, raising accuracy on unseen clips from 72% to 92% and cutting word substitutions by 66%.
Google
Feb 2026 - May 2026
Contract Software Engineer, Codebase Collaboration
Sunnyvale, CA
- Developed a Gemini/Vertex AI tool router spanning 5 tools for an NL2SQL platform on GKE/AlloyDB; built a 768-dim pgvector index over 1.08 million transactions, with SQL execution, embedding, and vector retrieval each under 550 ms.
C.Light Technologies
Aug 2025 - Jan 2026
Software Engineer Intern
Berkeley, CA
- Built a hybrid reasoning layer that converts ambiguous natural-language queries into structured outputs for multi-agent routing and execution.
- Applied language-model disambiguation and multi-turn context propagation to maintain conversational state and resolve follow-up queries.
Amazon
May 2025 - Aug 2025
Software Engineer Intern
Seattle, WA
- Designed and piloted a multi-agent workflow using Strands Agents with AWS Bedrock and MCP for ticket triage, summarization, anomaly detection, and root-cause analysis across 200+ tickets, saving 25+ engineer hours per quarter.
smartQED
Jul 2022 - Dec 2023
Software Engineer Intern
Remote
- Implemented a summarization pipeline leveraging the T5 transformer model to generate insights from 1,000+ Stack Overflow posts.
- Designed real-time analytics dashboard with Java/MySQL backend, implementing RegEx filtering to process 500+ DB entries.
- Trained a custom NER model using spaCy for domain-specific classification, improving summarization accuracy by 33%.
Education
University of California, Berkeley
Expected May 2028
B.S. Electrical Engineering and Computer Science, Minor in Data Science
Berkeley, CA
Leadership: Vice President, UC Berkeley Codebase; Staff, Agentic AI Summit 2026
Coursework: Artificial Intelligence, Machine Learning, Data Structures, Computer Architecture, Data Science Principles, Computer Simulations, Discrete Math & Probability Theory, Structure and Interpretation of Programming Languages, Linear Algebra, Multivariable Calculus.
Skills
Languages
ML / AI
Systems & Infrastructure
Projects
What I build when nobody tells me what to build.
No projects match. Try another topic or search.
Blog
Longer write-ups on projects, experiments, and what I learned from them.
Technical Report2026
ActionRank
Does Netflix's "score, don't generate" idea work for agent tool selection? Scoring a fixed tool list in one forward pass versus generating the tool name token by token, tested on ToolBench.
Read →
Research Note2026
Fooling a Trusted Monitor
A fine-tuned Laya monitor for agent traces is near chance on held-out sabotage, and a white-box attack flips its verdicts cheaply wherever the evidence is the agent's own words, sometimes by rewording what the agent claims it did.
Read →
Contact
Always happy to talk about hard problems.
LinkedIn
linkedin.com/in/yashdthapliyal
GitHub
github.com/YashDThapliyal