Ankit Dahal

Projects

  • Agentic data synthesis platform

    Current Agents · Orchestration · Knowledge graph

    Building a coordinated multi-agent setup using the Claude agent SDK to ingest thousands of API sources at scale with divergent schemas.

    • Orchestrator agents — coordinating search jobs and self-healing.
    • Search agents — autonomously find datasources and build ingestion pipelines.
    • Browser use agents — captcha solving and javascript execution.

    Millions of data points today, architected toward billions. Lineage, freshness, and cross-source joins power a single analytical surface for production-grade data synthesis.

  • Browser use agents

    VLM · Agents

    Built multiple browser use agents with frontier models. Includes a self-built stealth browser harness intended for captcha solving. Supports research tasks: answering questions and validating answers based on multi-document source material on the web. Frequently repurposed for any task where agents need access to websites.

  • Lambda Cloud MCP

    MCP · ML infra · Agents

    MCP surface for Lambda Cloud — auto-provision GPUs, orchestrate ML environments and training jobs, steer agents by text message. Code

  • Legal Agentic RAG System

    RAG · Agents

    Agentic RAG on official legal pages with various hybrid search implementations, reranking, multi-agent systems. 90% retrieval accuracy on evaluations created from high quality synthetic data. Finetuning on Google cloud platform with torch distributed elastic training. 10,000+ HuggingFace dataset downloads. Code · Training infra · Writeup

  • LoRA Targeting for Persona SFT

    Fine-tuning · MoE · VLM

    Gradient-guided LoRA targeting for persona SFT on qwen3.5-35b-a3b; Curated 700,000+ noisy discord messages into high quality multi-turn SFT data; lmms-eval OSS contribution (eval harness used by frontier labs & thousands of users). Accelerated evals by 30x with vLLM. Code · Writeup

  • Llama 3.1 8B Instruction Tuning

    Fine-tuning

    LoRA SFT with +52% IFEval improvement (200→305/834) on ~$10 compute, and multiple quantization experiments. Code

  • Large Language Model Pretraining

    Pretraining · DDP

    Pretrained a 450M-param transformer on FineWeb-Edu (10B tokens), 8xA100 gpus using DDP — RoPE, SwiGLU, Flash attention, GQA/MHA, KV-Cache, etc. Code · Checkpoint

  • Transformer from Scratch

    Fundamentals

    PyTorch "Attention Is All You Need" implementation with encoder-only, decoder-only, and encoder-decoder architectures. Along with modern extensions (SwiGLU, RMSNorm, RoPE) Code

Technologies

Applied AI
Python, AI Agents, Multi-agent systems, Retrieval augmented generation, Finetuning, Evaluations, Large language models (LLM/SLM), Multi-modal models, Vision language models, Claude Agent SDK, Playwright
Applied ML
PyTorch, HuggingFace, Transformers, Pretraining, Supervised finetuning, LoRA/PEFT, Quantization, Unsloth, Bitsandbytes, Weights & Biases, Torch Elastic, Distributed Data Parallel (DDP)
ML infra
Docker, Google Cloud, Google Compute Engine, Google Kubernetes Engine, Lambda Cloud, vLLM, Vector databases, Hierarchical Navigable Small World (HNSW) tuning, ElasticSearch

Writing

Technical Writeup

LoRA Targeting for Discord-style SFT

Deep-dive on gradient attribution-guided LoRA targeting, dataset curation from raw Discord exports, training/eval loops, and checkpoint selection for authentic persona style transfer without catastrophic forgetting.
Technical Writeup

Agentic RAG with production level accuracy

End-to-end design notes covering hybrid retrieval architecture, embedding strategies, reranker integration, agentic orchestration, and lessons learned building a RAG pipeline on real legal documents with production level accuracy.