Skills
What Suryansh Chaudhary builds with
Not a checklist — every tool below is here because something shipped with it. Where a claim maps to a project or a build log, the link is on the page.
AI agents & LLM engineering
The core of the day job: multi-agent research systems that have to survive contact with real users and real money. Explicit state, small nodes, narrow tool boundaries, and gates on the decisions a model should not be making.
- LangGraph
- Control plane for a multi-agent financial research desk — conditional routing, explicit graph state, and supervisor nodes that decide which specialist runs next.
- LangChain
- Tool integrations and model adapters around the graph, including cost-aware routing between model tiers for different nodes.
- Agentic RAG
- Retrieval that plans, checks its own evidence, and only then answers — built on Vertex AI's managed multimodal RAG Engine so charts and tables in filings survive ingestion.
- Prompt-injection defence
- Layered guardrails: deterministic hard gates for the non-negotiables plus an LLM-as-judge soft gate for ambiguity, after finding the refusal message itself was leaking the system prompt.
- Gemini API
- Multimodal generation and image pipelines, including the script that generates every cover image on this blog.
- OpenAI API
- Streaming completions and tool calling behind a provider-agnostic interface.
- Model Context Protocol (MCP)
- Server and payload design — closed vocabularies and schema-constrained returns, because the payload is the real integration surface.
- Vector search & embeddings
- Chunking strategy, retrieval evaluation, and groundedness scoring on financial documents.
Backend & APIs
Treating the model as an unreliable, expensive network call, and building everything around it to be fast, cheap and boring.
- Python
- Primary backend language for agent services and data pipelines.
- FastAPI
- Streaming LLM responses, background jobs, and the API layer the agent frontend talks to.
- Django
- Conventional CRUD products where batteries-included wins.
- Node.js
- CLI tooling and realtime services — Signal Boy is a zero-dependency Node CLI with per-OS radio backends.
- Socket.IO
- Oratora's realtime tier, scaled separately from the app and load-tested at 1,000 simultaneous joins.
- REST & GraphQL
- Versioned APIs, typed SDKs, and error contracts that clients can rely on.
Frontend
Interfaces for systems that stream, fail partway, and change their answer mid-render.
- React
- Component architecture and state management for agent-driven UIs.
- Next.js
- App Router, server components, route-level metadata and static generation — this site and Oratora both run on it.
- TypeScript
- Strict types end to end, including shared contracts between backend and client.
- Tailwind CSS
- Token-driven design systems rather than ad-hoc utility soup.
- Framer Motion
- Scroll-linked and gesture animation that respects reduced-motion.
macOS & Swift
MacGet is the proving ground — a native download manager where concurrency bugs show up as a stalled progress bar rather than a stack trace.
- Swift
- An actor-based download engine handling chunked parallel HTTP transfers.
- SwiftUI
- The full MacGet interface, including queue management and live transfer state.
- Swift concurrency
- Actors and structured concurrency to keep chunk writers, the UI, and the scheduler from fighting over shared state.
- Native Messaging
- Browser-extension capture that hands downloads from Chrome and Safari to the app.
- BitTorrent & yt-dlp
- Torrent transfers and bundled yt-dlp for one-click video downloads.
Data
Storage picked for the access pattern rather than the résumé.
- PostgreSQL
- Relational core for Oratora and the blog's subscriber and likes data.
- MongoDB
- Document storage for semi-structured agent traces and run history.
- Redis
- Pub/sub fan-out for realtime, plus caching and rate limiting in front of expensive model calls.
Cloud & DevOps
Shipping and paying for AI workloads — the part that decides whether a demo becomes a product.
- Google Cloud Run
- Containerised agent services with request-scoped scaling.
- Vertex AI
- Managed multimodal RAG Engine as the data plane under LangGraph.
- Docker
- Reproducible builds for Python services and CI.
- GitHub Actions
- CI/CD, release automation, and macOS build and notarisation for MacGet.
- Vercel
- Edge hosting for this site and Oratora's web tier.
- Linux
- Day-to-day server work and per-OS backends for Signal Boy.
Working together
Open to collaborations and freelance work on AI agent systems, agentic RAG pipelines, and full-stack products around them. The projects and build log are the honest version of this page — including the decisions that turned out wrong.