Harshit Dave

I am a Senior Data Scientist at slice, where I build production AI systems powered by large language models, multimodal learning, and distributed machine learning. My work spans multi-agent reasoning systems, LLM fine-tuning, document intelligence, conversational AI, and speech interfaces, with a focus on translating cutting-edge AI research into reliable production systems.

My journey in AI/ML began at IIITM Gwalior, where I reseached and engineered the Integrated-Particle Swarm Optimization (i-PSO) technique and analyzed with various ML algorithms for detecting malware, under the supervision of Dr. Saumya Bhadauria and Prof. Aditya Trivedi. This early work taught me how to blend deep theoretical concepts with practical, real-world problem-solving.

Later, as a remote research intern with the artificial intelligence institute of University of South Carolina (AIISC), I worked with a team on multimodal fact verification. We used Stable Diffusion 2 to enhance the mutlimodal data and built more robust datasets (FACTIFY3M) by injecting adversarial attacks in the form of fake news. This research, which was accepted at EMNLP 2023, gave me firsthand experience in working on multimodal AI.

During my internship at IBM Research, I benchmarked open-source LLMs for enterprise tasks and developed a 3-stage LLM-based approach for NBA (next best agent) recommendation in cold-start settings, work that was accepted at IAAI-24 (co-located with AAAI-24).

My Work at slice

At slice, I work on building production AI systems spanning fraud detection, conversational AI, voice interfaces, document intelligence, and LLM training. My work focuses on taking modern AI research into production through agentic architectures, multimodal models, distributed training, and scalable inference systems.

AI Fraud Investigation Co-Pilot: Architected a production multi-agent LLM system that aggregates onboarding, authentication, and transaction signals into structured investigative summaries for fraud analysts. Built the agent orchestration, reasoning pipeline, evaluation framework, and production inference stack to enable faster and more consistent fraud investigations.

Decision Intelligence: Fine-tuned a DeBERTa-v3 classifier on investigator-labelled fraud cases using DeepSpeed across multi-GPU infrastructure. Developed the complete training, evaluation, and deployment pipeline for production-scale decision support using modern transformer fine-tuning techniques.

Document AI: Built an end-to-end document understanding pipeline using YOLOv8, OCR, and domain-specific parsing heuristics to extract structured information from identity documents. The system replaced an external OCR dependency while improving operational flexibility and significantly reducing inference costs.

Voice AI: Engineered a multilingual voice payment assistant combining speech recognition, LLM reasoning, tool execution, and policy-based guardrails for low-latency conversational payment workflows.

Conversational AI: Designed an agentic customer-support framework powered by GPT-class models with retrieval, conversation memory, workflow orchestration, and evaluation pipelines, enabling high-quality automated customer interactions in production.

AI Safety & Mechanistic Interpretability

In 2026, I was selected as one of 40 participants from over 1,500 applicants for the International Programme on AI Evaluation: Capabilities & Safety , a fully funded international programme focused on AI evaluation, mechanistic interpretability, alignment, and AI safety. The programme brings together researchers and engineers to study rigorous evaluation methodologies for frontier AI systems under mentorship from experts across academia, industry, and AI safety organizations.

During the programme, I investigated narrative-driven emergent misalignment by fine-tuning Qwen3-4B on synthetic persona data and analyzing how localized training signals propagate into broader behavioral changes. Using linear probes, activation analysis, and layer-wise interventions, I studied how representations related to personality, morality, and safety emerge inside transformer models and how robust they remain after behavioral unlearning.

Research Interests

I am interested in developing AI systems capable in symbolic reasoning and build efficient collaborative systems that works alongside people in meaningful ways. My current focus is on building collaborative, agentic models that can understand goals, share context, and adapt through interaction, ideas that connect closely with Human-Computer Interaction (HCI) and Human-Centered AI (HCAI).

Currently Exploring

• Exploring how reinforcement learning can shape LLM personas for collaborative agent systems, specifically through post-training a Qwen3 MoE model
• Currently digging into this paper, which takes an interesting approach—instead of relying on expert-annotated feedback, it uses real human interactions to develop the model's persona
• Experimenting with the Muon Optimizer, which has been showing better training efficiency than traditional optimizers in recent benchmarks
• On the implementation side, diving into a source code that has claimed to have improved performance over flash-attention to see where I can squeeze out performance gains
• Learning Triton along the way since the performance optimization code is written in it