Ankit Dahal

About

I took the decision to leave my role in 2025 to create a path for myself to pivot into AI/ML engineering. Years of foundational AI knowledge have been compressed into months of strenuous study and application through personal projects. This has allowed me to move exponentially faster than I would have been able to otherwise, building a strong understanding of modern AI from first principles.

I am looking for technically challenging problems that matter. I strive to be a strong generalist, diving deep when needed, and am passionate about a wide range of topics from AI, distributed systems, and the sciences.

Projects

Browser use agents & harnessing

In progress Current AI Safety

Mapping AI-enabled manipulation across major social media platforms. Studying coordinated inauthentic behavior at scale.

View details

Mapping the threat landscape of AI-enabled manipulation across major social media platforms. Studying coordinated inauthentic behavior: how bot-driven narratives (astroturfing) exploit platform mechanics and cognitive biases at scale.

5+ Platforms
AI + Psychology + Cybersecurity Interdisciplinary

Technology: Qwen3.5-397B-A17B (multi-modal/VLM), playwright browser harness, sft & reinforcement learning, python

Adversarial AI Coordinated agents Platform analysis Cognitive psychology AI safety
Project Outline →

Large Language Model Pretraining

MVP Pretraining

450M parameter transformer inspired by the gpt and llama papers, pretrained on 10B tokens using 8xA100 GPUs.

View details

End-to-end pretraining of a custom, self-built 450M parameter GPT/Llama-style transformer on FineWeb-Edu (10B tokens). Trained on 8xA100 GPUs using Distributed Data Parallel with custom training loops and memory optimizations.

450M Parameters
10B Tokens
8x A100 GPUs
PyTorch, DDP Custom training loop Chinchilla scaling law RoPE, RMSNorm, SwiGLU, GQA/MHA, KV cache Flash attention
Code → Model Checkpoint →

Legal RAG System

MVP RAG / Agents

Production-grade agentic RAG for legal documents with hybrid dense + sparse retrieval.

View details

Production-grade agentic RAG system for legal documents. Full pipeline from data sourcing through retrieval, reranking, and agent orchestration. Hybrid search combining dense embeddings (bge-m3) with sparse retrieval (BM25).

bge-m3 ChromaDB Elasticsearch bge-reranker Gemini-2.5 Docker GKE

Embeddings: bge-m3, gemini-embedding-001

Indexing: ChromaDB (HNSW), Elasticsearch (BM25)

Retrieval: Hybrid search with RRF, convex combination

Agents: Conversational + search agents, planning, self-triage

Training: 560M parameter model fine-tuning

Code → Training Infrastructure →

Llama 3.1 8B Instruction Tuning

Fine-tuning

LoRA fine-tuning with +52% IFEval improvement using ~$10 compute.

Details

LoRA fine-tuning of Llama 3.1 8B for instruction following. Achieved +52% IFEval improvement (200 → 305/834) with ~$10 of compute across 4-bit, 8-bit, and BF16 quantization experiments.

LoRA Unsloth Quantization IFEval W&B
Code →

Transformer from Scratch

Fundamentals

Clean PyTorch implementation with modern LLM improvements.

Details

Clean PyTorch implementation of "Attention Is All You Need," extended with modern LLM architectural improvements: RoPE, SwiGLU, RMSNorm.

PyTorch MHA SwiGLU/ReLU RMSNorm
Code →

Writing

Technical Writeup

Building a Production-Grade Legal RAG System

End-to-end design notes covering hybrid retrieval architecture, embedding strategies, reranker integration, agentic orchestration, and lessons learned building a RAG pipeline for legal documents.

Open to Opportunities

Looking to join teams building interesting AI systems.

View Resume GitHub LinkedIn