Fine-tuning and optimizing LLM-powered agents with GPT-4o, Gemini, and Claude, cutting hallucinations, tightening response consistency, and wiring external tools into enterprise finance workflows.
AI · ML · Research
Daksh Agarwal
Building AI that thinks carefully and ships cleanly.
Machine learning engineer & researcher: LLM agents, RAG systems, computer vision, and evaluation pipelines for real-world systems.
Move to explore · Scroll to continue
About
Dual-degree AI student crafting production agents, research benchmarks, and vision systems that hold up outside the lab.
Currently a Machine Learning Intern at Peakflo (YC W22) and a Research Associate at NYU. I work across LLM agents, predictive maintenance, RAG evaluation, and object detection, with a bias toward measurable gains and responsible AI. Previously at Avawatz, Grant Thornton, and HCL Guvi.
Pursuing a B.Tech in AI & ML at GGSIPU and an Online Degree (B.S.) in AI & Data Science at IIT Madras. Published at ASIACCS (A) and in MDPI Electronics (Q2). ML Lead at Google Developer Group.
Experience
Where I build
Co-leading RAG benchmarks vs Open Deep Research and AutoGPT across retrieval quality, grounding, and hallucination rate. Multi-LLM pipelines with GPT-4o, Gemini, Llama, and Claude, with +9% answer accuracy via prompt and retrieval optimization.
Enhanced GenMobius enterprise AI MVP; intent-aware agents with Azure AI Search; proposal generation system; NASA C-MAPSS RUL predictive maintenance benchmarks.
Predictive failure detection for Flint bus systems; 100k+ synthetic records for rare events; Graph Neural Networks outperforming classical and deep baselines. Also fine-tuned YOLOv8 on 800+ dental radiographs to 91% mAP50.
Designed 10+ cloud & AI curriculum modules. Co-built an AI avatar & synthetic voice MVP (HeyGen-style) with generative APIs and audio synthesis in a deployable web app.
Built a real-time underwater object detection system with YOLOv8 on a custom dataset from proprietary camera footage and open-source data. Validated on physical hardware in live underwater environments; designed the full ingestion, annotation, and augmentation pipeline.
End-to-end deep learning pipeline for malware classification from opcode and API sequences across 20+ families. Decompiled 10k+ Windows binaries with IDA & Ghidra; transformer NLP reached 96.4% accuracy with SHAP-based interpretability.
Projects
Selected builds
Prompt Sensitivity in LLM Code Generation
~6,000 evaluation jobs across six LLMs on Azure AI Foundry. Custom metrics (PromptVar, WorstPrompt, SensRange) revealing robustness gaps that pass@k misses.
Domain-Specific RAG Chatbot
LangChain + vector store with scoped retrieval, session memory, and prompt boundaries for governable, on-topic multi-turn answers.
Underwater Object Detection
Custom YOLOv8 pipeline from proprietary + open footage through annotation, augmentation, and live hardware validation in real underwater conditions.
Malware Classification Pipeline
Opcode and API sequences from 20+ families; transformer NLP on 10k+ Windows binaries decompiled with IDA & Ghidra: 96.4% accuracy with SHAP audits.
Skills
Stack
- Languages
- Python, C, C++, SQL, JavaScript, HTML/CSS, R, Bash
- Machine Learning
- PyTorch, TensorFlow, Scikit-learn, XGBoost, LightGBM, CatBoost, FastAI
- LLM / NLP
- LangChain, LangGraph, LlamaIndex, RAG, Transformers, Prompt Engineering
- Vision & MLOps
- OpenCV, YOLO, Docker, Kubernetes, MLflow, FastAPI, Flask
- Cloud
- Azure ML, Azure AI Foundry, Azure OpenAI, GCP Vertex AI, OCI
- Vector Stores
- FAISS, Chroma, Pinecone
Education
Study & pubs
Guru Gobind Indraprastha University
B.Tech, Artificial Intelligence & Machine Learning · CGPA 8.0
IIT Madras
Online Degree, B.S. in Artificial Intelligence & Data Science · CGPA 8.0
Leadership
ML Lead, Google Developer Group
Publications
- DECKER: Domain-invariant Embedding for Cross-Keyboard Extraction and Recognition ACM ASIACCS 2026 (A Ranking)
- CustomNerd: A Framework and Tool for Fast Deployment of Production-Ready Expertise-Based Question Answering Systems MDPI Electronics 2026 (Q2)
Contact
Let’s build something sharp.
Open to research collaborations, ML internships, and agentic systems work.