Kshitiz Waglebackend · AI systems · agents
Open to backend, platform, and AI engineering rolesSenior Software Engineer · Rippey AILalitpur, Nepal (UTC+5:45)Professional VibeCoder

I build backends for document and AI workflows.

Async Python services, message-driven pipelines, document processing, and multi-agent orchestration. I started as a Process Analyst and moved into engineering largely self-taught, so I start from the operational problem, then build the system that removes it.

Lately on GitHub
Five commits were applied, documentation was tightened around ScaleDown templates and it was clarified that a blank system prompt disables them. The per‑model sampling allow‑list was removed, time_log was added to version‑2 response metadata, and a third‑generation structure model was baked into the image by the layout update. Alongside a couple of pull requests being closed, the changes streamline AI orchestration and packaging without altering external interfaces.
refreshed hourly from GitHub
Since
Jun 2022 · four roles at Rippey AI
Core stack
Python · FastAPI · RabbitMQ
Looking ahead
ML security and AI safety, where the agent infrastructure I build meets the question of keeping it trustworthy
The past twelve months · one crest per week

268 active days · best streak 20 days · weekends mostly off

4+ yrsShipping at Rippey AI since Jun 2022
4 rolesProcess Analyst → Senior Software Engineer
PythonFastAPI · RabbitMQ · Redis · MongoDB
UTC+5:45Remote, from Lalitpur, Nepal

01 — Work

Work I can show

What I shipped most recently, the four systems underneath it at Rippey AI, then public projects you can explore. Employer code is private, so this describes architecture and decisions rather than code.

Shipped · agent infrastructure

Synapse

A containerized agent harness built on Claude Code that enforces skill-scoped execution over WebSocket. It is finished, and it was the center of my recent work.

  • Claude Code
  • WebSocket
  • Redis
  • Docker Compose
  • Agent architecture
Hierarchical orchestration
Restructured a monolithic agent into layers: an orchestrator delegates to a workflow builder, which coordinates specialist agents for ambiguity resolution, node construction, graph validation, and output formatting.
Auth, designed end to end
Designed the onboarding auth flow with session binding and per-turn authorization checks in Redis, documented as sequence diagrams.
Cost-aware model routing
Routed subagents by workload after comparing Sonnet, Gemini Flash, and Haiku pricing against what the agents actually do.
A hardened runtime
Fixed gunicorn permission errors, Docker Compose networking issues, and ambient configuration leaking into agent subprocesses.
Ongoing

Claude Code tooling

Agent behaviors turned into reusable skills, a skill-creator meta-skill, a diagrammer skill, plus hooks, MCP servers, and subagent routing tuned for long sessions.

  • Claude Code
  • MCP
  • CI
Planned

Resonance

A music recommendation app with a multi-provider LLM fallback chain, so recommendations keep working when one provider fails.

  • React
  • FastAPI
  • PostgreSQL

Foundations at Rippey AI

Document processing

01

The problemPDFs, email, scans, Office documents, and structured data arrive with different extraction and conversion needs.

What I didI built the Python services that extract, OCR, normalize, and route them, with fallbacks for the attachments, scans, and Office files that break the first path.

  • Python
  • FastAPI
  • OCR

Async services

02

The problemDocument work moves between services, and a worker can disappear halfway through a job.

What I didMessages are typed and acknowledged, consumers start and stop with the service, and retries and reconnects let the work resume.

  • FastStream
  • RabbitMQ

Workflow orchestration

03

The problemEach document workflow needs its own model behavior, and its state has to agree across services.

What I didI moved prompts, models, and workflow templates into validated configuration, and kept workflow state in sync across the stores services share.

  • LLM APIs
  • Redis
  • MongoDB

Product interface

04

The problemLive document workflows need to preserve review and session state across reconnects and reloads.

What I didI vibe-coded the React and Next.js interface over WebSockets, including live chat and session restore after reconnects.

  • React
  • Next.js
  • TypeScript
  • WebSockets

Public projects

Browse every project →

02 — Experience

From process analyst to senior engineer

Four roles at Rippey AI, newest first, then the degree. The line under each role is that stretch of shipping, week by week.

  1. Dec 2024 → nowRippey AI · Remote
    Senior Software EngineerCurrent

    Shipped Synapse and build the document services around it, from message handling to model configuration.

    Scope and focus

    Scope: Build the services behind document workflows, and built Synapse, a multi-agent orchestration system, plus the messaging, model configuration, and workflow state around it.

    Focus: Most of the work sits at the boundaries: queues, schemas, service lifecycles, and recovery when something upstream goes away.

  2. Aug 2023 → Dec 2024Rippey AI · Remote
    Software Engineer

    Built the services that classify and extract business documents for downstream systems.

    Scope and focus

    Scope: Built Python document services that classify, extract, and shape data for downstream systems.

    Focus: Worked on model serving inside async APIs, document-splitting edge cases, and integration data contracts.

  3. May 2023 → Aug 2023Rippey AI · Remote
    Automation Analyst

    Automated the handoffs: file movement, notifications, spreadsheets.

    Scope and focus

    Scope: Automated recurring operational handoffs: file movement, notifications, and spreadsheet reporting.

    Focus: Also handled the document exceptions that break those handoffs, such as scans that split badly.

  4. Jun 2022 → May 2023Rippey AI · Remote
    Process Analyst

    Reviewed logistics and accounts-payable documents, then turned the recurring checks into scripts.

    Scope and focus

    Scope: Reviewed logistics and accounts-payable documents, then scripted the checks that did not need judgment.

    Focus: Turned repeated manual review into classification, extraction, and reporting scripts.

  5. 2019–2023Academia International College · Tribhuvan University
    Bachelor of Computer Application

    Programming, databases, and networking fundamentals. I finished the degree alongside work.

    Scope and focus

    Scope: Before starting work in June 2022, I was a BCA student learning programming and computer-science fundamentals. I continued the degree alongside work through 2023; student coursework includes SQL, programming exercises, Java networking, Android screens, and Python API projects, shown separately from employer services.

    Focus: Coursework spans SQL, Java networking, Android, Python APIs, and the usual C-family foundations.

03 — Stack

What I reach for

Languages
  • Python
  • TypeScript
  • JavaScript
  • Rust
Student foundations
  • Java
  • Android
  • C
  • C++
  • C#/.NET
  • PHP
Frameworks
  • FastAPI
  • Pydantic
  • FastStream
  • asyncio
  • Flask
Data & messaging
  • RabbitMQ
  • aio-pika
  • Redis
  • PostgreSQL
  • MongoDB
  • AWS S3
  • SQL
AI & agents
  • Multi-agent orchestration
  • Synapse
  • Claude Code
  • MCP
  • LangChain
  • Gemini
  • LLM APIs
  • Structured output
Infrastructure
  • Docker
  • GitHub Actions
  • pre-commit
  • Cue
  • Cloudflare Workers
  • Cloudflare Tunnel
  • Caddy
  • Authelia
  • Bash
  • PowerShell
  • SSH
Document & OCR
  • Tesseract
  • Azure OCR
  • pikepdf
  • Ghostscript
  • OpenCV
wrote about it →
Web & reporting
  • React
  • Next.js
  • Astro
  • Vite
  • WebSockets
  • pandas
  • Streamlit

04 — Field notes

Notes

Short notes about PDFs, Office files, OCR, and the small format decisions that turn a clean pipeline into a debugging session.

All notes →

05 — Off the clock

When the laptop closes

Rides

Roads on a scooty

Rides out of Lalitpur, traced along the roads taken. 2,040 km across 7 so far.

homeLumbini
Home lab

One cracked laptop, several services

An old laptop lost its screen and gained a second career as an Ubuntu Server box. Jellyfin, Sonarr, Radarr, Immich, Navidrome, and Caddy sit behind a Cloudflare Tunnel with Authelia. The best save so far: recovering the Immich PostgreSQL database from WAL corruption. Configuration still takes longer than using any of it.

Coffee & cafes

Cold brew, americano, pour-over

Cold brew when it's hot, americano when I want something warm, pour-over when I'm not in a rush. Dark roast wins. Lighter roasts are too acidic and my stomach already has opinions.

Games

A reliable escape hatch

RPGs for switching off. Genshin when I want something to do with my hands. Fate/Grand Order for the story, which is a strange sentence until you meet its novel-length lore across multiple singularities. An open world or a long story reliably clears the mental cache. Most recently finished: a first playthrough of Black Myth: Wukong. Phantom Blade Zero is next.

06 — Contact

Hiring for backend, platform, or AI engineering?

Email is the best way to reach me. Share the role and what the team is building.

New messageLalitpur
Subject
Role: [title] at [company]

Hi Kshitiz,

[What the team is building]

[The role, and where it sits]

[Remote, location, and timeline]

Open in mail app