Software Developer
Arnav Panigrahi
Backend systems, full-stack applications, and AI/ML pipelines shipped to production.
Background
About
I write backend services and ML systems that run in production. Lately that's been FastAPI microservices, LangChain agents, and multi-modal models. I did my M.S. in CS at UC Riverside and I'm based in California.
Outside of work, I play guitar, sink hours into JRPGs, and tinker with OSS and agentic harnesses. I collect keyboards — HHKB is the only correct answer — and use Mario Zechner's pi agentic harness. I'm slowly teaching myself Japanese.
Experience
Where I've worked.
Software Developer
AllCheer- Built a LangChain AI agent for ROI workflows, automating authorization checks, compliance validation, and document generation while cutting processing time by 60% and saving $20K/year.
- Engineered a containerized FastAPI service that emulates authenticated RethinkBH sessions for real-time clinical record sync, bypassing 24-48-hour public API latency.
- Developed a React + FastAPI NER pipeline using OpenAI for structured data extraction from therapist notes at 92% accuracy.
Software Developer Intern
AllCheer- Designed a Python logistics engine on Google Cloud with Google Maps APIs for multi-stop route optimization and large-scale staff scheduling.
- Led backend R&D and API schema design for ROI automation, moving Make.com prototypes into production service architecture.
Junior Software Developer
Pinnacle Consulting LLC- Integrated ArcGIS Online with Angular and .NET for a $75M utility provider, enabling real-time infrastructure visualization and reducing oversight errors by 30%.
- Delivered three full-stack ERP systems with Node.js and SQL Server, implementing JWT-based RBAC and complex query workflows to automate invoicing.
Capabilities
Tools I work with.
Languages & Frameworks
- Python
- TypeScript
- JavaScript
- C#
- .NET
- FastAPI
- React
- Angular
- Astro
AI & Data Systems
- LangChain
- LangGraph
- PyTorch
- Vector Search
- Embeddings
- Prompt Engineering
Infrastructure & Delivery
- Google Cloud
- Docker
- AWS
- REST APIs
- Node.js
- SQL Server
- PostgreSQL
- SQL
- Git
Selected Work
Projects
bedtime.ai
Multi-Modal AI Storytelling Platform
- 4 models orchestrated in a single pipeline
- ~70% classification accuracy on children's drawings
- 35 backend + 19 frontend tests
Sabi Agent
Rust Coding-Agent Harness
- OpenAI-compatible agent loop with read, write, edit, bash, search, and web tools
- JSONL sessions with cwd-aware resume, command history, and skill invocation
- Tauri desktop shell with project selection, approval cards, and collapsible diffs
PR Review Agent
Cloudflare GitHub App Review Bot
- Webhook-first PR review flow with HMAC verification before parsing
- Bounded diff and file-context collection with typed Valibot review output
- Update-in-place review comments with severity, recommendation, findings, and metadata
Yomiji
Offline-First WaniKani Android Client
- SQLite incremental sync with pending-write queues
- Romaji-to-kana answer checking, offline-first
- Jest-covered domain logic for lessons and reviews
Other Projects
A real-time collaborative site with live cursor sharing and viewer presence. Built on Cloudflare Workers, Durable Objects, and SvelteKit 5.
View on GitHubAn interpretability study on LiquidAI/LFM2.5 that identifies and removes refusal directions via weight orthogonalization. Reduced model refusal rate from 50% to ~37% without retraining.
View on GitHubA containerized FastAPI service deployed on Google Cloud Run that emulates authenticated user sessions against RethinkBH endpoints to sync appointment and authorization data in real time, bypassing 24-48-hour API latency for downstream applications.
View on GitHubA production-ready RAG pipeline for querying Star Wars scripts with LangChain, OpenAI embeddings, and Qdrant. Expanded from a tutorial into a configurable system with prompt-injection hardening, source attribution, ~30x cost reduction using GPT-4o-mini, and a 28-test suite.
View on GitHub