AI engineer and full stack developer

LLM systems
that hold up
in production.

I build agents, guardrails, evals, and computer vision pipelines, then sit with the people using them until the thing actually works. Most of my time goes into the part after the demo.

Portrait of Suyash Bhattarai
  • World 2nd in CS Official Cambridge Recognition
  • Top in country in CS Outstanding Cambridge Awards
  • 1540 SAT 99th percentile globally
  • A*s in CS and Maths Cambridge A Levels
  • Best Academic Performance Award Budhanilkantha School
  • CAIE AS Level Topper Budhanilkantha School
  • As in Econ, Business and EGP Cambridge A Levels

Selected work

Agents, infrastructure, and applied ML.

Open Switch

LLM gateway with circuit breakers, half-open probes, hedged requests, cost-aware routing, per-key budgets, and request traces.

  • Rust
  • Tokio
  • Axum
  • WASM

Screener

Typo tolerant hybrid search with faceted filters that runs fully in the browser.

  • Rust
  • BM25
  • FST
  • WASM

Evalgate

LLM evaluation runner with its own YAML check format, flakiness detection, and a pull request gate that writes a Markdown report.

  • Rust
  • CLI
  • YAML
  • WASM

Corridor FX

FX exposure and hedging engine with a quantile forecast that is honest about what it cannot predict.

  • Python
  • XGBoost
  • FastAPI
  • React
05 Case study

Flamide

AI enquiry and CRM system in beta; collects student details, updates records and follows up with counselor handoff.

  • Next.js
  • TypeScript
  • Supabase
  • DeepSeek
06 Case study

Entode

Backend generator with a schema compiler, scoped model edits and a bounded repair loop driven by executable checks.

  • Python
  • FastAPI
  • pytest
  • React Flow

FX Risk Desk

Compare FX hedge policies with rolling VaR and expected shortfall, quantile forecasts, and option repricing on public reference rates.

  • Python
  • XGBoost
  • FastAPI
  • React

How I work

A green test suite is a claim, not a proof.

01

Run it live

My worst bugs passed every mocked test. I run the real provider, unintercepted, before I call anything done.

02

Guardrails outside the model

If a rule matters, code checks it after generation. Prompts ask nicely. Validators refuse.

03

Say what wasn't measured

Every report I write lists what it can't claim. A 0.55 AUC gets reported as a coin flip, not rounded up.

04

Sit with the users

I've shipped to factory floors, school offices, and police briefings. The spec is whatever they actually need on day two.

Contact

Let's build something
that holds up.