DA/26
loading field notes…
DA/26 résumé ↗
FIELD NOTES VOL. 03 KATHMANDU / NEPAL

currently experimenting with machines|

I make software
that thinks.

I'm Dev — an AI engineer and full-stack developer based in Kathmandu. I build agent systems that can research, reason, and take reliable action.

inspect my work ↓ say hello ↗
0 public repositories
0 agent systems
0 PyPI package
SPECIMEN 026 OBSERVATION
Dev Adhikari + +
SUBJECT DEV ADHIKARI
DOMAIN AI / SOFTWARE
BEHAVIOR BUILDS AGENTS
STATUS ACTIVE
↳ appears to be unusually
interested in transformers
MCP◆ COMPUTER USE◆ LANGGRAPH◆ PYTHON◆ UI AUTOMATION◆ OCR◆ VISION◆ AGENTS◆ FASTAPI◆ RAG◆ PYTORCH◆ TRANSFORMERS◆
01 OBSERVATION

Most of my work starts with a question rather than a technology.

What happens if an agent can use a computer reliably? Can a system observe, act, verify, and recover without pretending failure did not happen?

NOTE 01

I don't particularly enjoy building another todo app just because someone said it scales.

CURRENTLY BUILDING

Reliable computer-use agents

— observe, act, verify, recover
CURRENTLY EXPLORING

MCP tools and desktop automation

— UIA, AT-SPI, OCR, vision
CURRENTLY THINKING ABOUT

How agents earn user trust

— safety, verification, recovery
02 ARTIFACTS / SELECTED WORK
01
MCP / COMPUTER USE

Universal Computer Control

inspect ↗
OBSERVE
↓
UNDERSTAND
↓
ACT
↓
VERIFY
↓
RECOVER

A cross-platform MCP server that gives AI agents dependable control of Windows and Linux desktops through accessibility APIs, OCR, vision, and physical-input fallbacks.

↳ an agent should verify that an action worked, not just assume it did.

MCP PYTHON COMPUTER USE OCR / VISION
02
AI ASSISTANT / IN DEVELOPMENT

Sarah

inspect ↗
$ understand(task)
$ browse(web)
$ manage(files)
$ write(code)

→ execute(workflow)

An AI assistant designed to take action on a computer: browse the web, manage files, write code, and automate multi-step workflows instead of stopping at advice.

PYTHON AGENTS AUTOMATION
03
AUTONOMOUS NEWSLETTER AGENT

Vanguard

inspect ↗
VANGUARD research
↓
write
↓
schedule
↓
deliver

A long-running agent that researches a custom topic, produces a cited HTML newsletter, delivers it to a mailing list, and schedules follow-ups for time-sensitive events.

↳ useful autonomy means remembering what needs to happen next.

LANGGRAPH MCP PLAYWRIGHT PYTHON
04
RESEARCH ORCHESTRATION

Deep Research Workflow

inspect ↗
GATHER
↓
SYNTHESIZE
↓
PLAN
↓
WRITE
↓
REDUCE

A LangGraph workflow that turns a topic into a structured, citation-heavy report. It gathers context, plans sections, sends research workers out in parallel, then assembles the final brief.

LANGGRAPH TAVILY PYDANTIC OLLAMA
05
FROM FIRST PRINCIPLES

Transformer From Scratch

inspect ↗
I
am
learning
ATTENTION / 12 HEADS

A complete Transformer implementation based on Attention Is All You Need, trained for English → Hindi translation.

↳ built to understand what the abstraction was hiding.

PYTORCH NLP TRANSFORMERS
06
LEGAL / RETRIEVAL

CCPA Compliance RAG

inspect ↗
EVIDENCE Qwen2.5-1.5B
↓
ChromaDB
↓
semantic retrieval
↓
FastAPI

A lightweight retrieval-augmented system for analyzing potential CCPA violations, grounded in the relevant statute and exposed through a FastAPI endpoint.

RAG CHROMADB FASTAPI QWEN
03 FIELD NOTES
NOTE / 001

The feedback loop is the product.

For computer-use agents, acting is only half the job. The system has to observe what changed and know when to try a different path.

computer use
NOTE / 002

Fallbacks are a feature.

Good automation does not depend on one perfect interface. Accessibility data, OCR, vision, and input controls each cover a different failure.

MCP
NOTE / 003

Autonomy needs memory.

Long-running agents become useful when they can preserve context, schedule the next step, and report what they did along the way.

orchestration
04 CHRONOLOGY
FIELD

Agent Infrastructure

Building MCP tooling and computer-use systems that help AI agents observe, act, verify outcomes, and recover from failures.

MCP Python Computer Use OCR / Vision
FIELD

AI Agent Development

Research workflows, scheduled newsletter automation, and tool-using systems built with explicit orchestration and structured outputs.

LangGraph Playwright Pydantic FastAPI
FIELD

ML / Deep Learning

Hands-on work with Transformers, BERT, ViT, and retrieval systems across language, vision, compliance, and financial datasets.

PyTorch Transformers ChromaDB Qwen
FINAL OBSERVATION

Have an interesting
problem?

I'm more interested in strange problems than impressive job titles. If you're building something that makes you think, I'd like to hear about it.

send a signal ↗
github ↗ linkedin ↗