I make LLM features dependable enough to ship.

Switch between OpenAI, Gemini and AWS Bedrock with one environment variable and zero code changes.

```json
{"criterion": "B2",
 "awarded": "2",
 "follow_through": true}
```

Here is the result:
{"criterion": "B2",
 "awarded": 2,
 "followThrough": "yes"}

{"criterion": "B2",
 "awarded": 2,
 "follow_through": true,
 "note": "method mark"}

{"criterion": "B2",
 "awarded": 2,
 "follow_through": true}

Problem

Code that calls one LLM vendor directly ties its cost and reliability to that vendor, and vendor pricing and capabilities keep changing.

What I built

One interface over all three providers. Call sites ask for a result and never name a vendor, and every response is normalised into one shape.

Key decision

Send non-critical tasks to cheaper models behind the same interface.

• Python
• OpenAI
• Google Gemini
• AWS Bedrock Converse API

Grades answer sheets for 3+ exam boards and replaces hours of manual marking per batch.

Problem

Teachers spent hours per batch applying mark schemes by hand, and each exam board formats its schemes differently.

What I built

Parses question papers and mark schemes straight from PDFs, then grades answer sheets against per-criterion rubrics.

Key decision

Follow-through logic, so one early mistake doesn't cost marks twice.

• Python
• FastAPI
• PDF parsing
• LLM-based evaluation

Two real-time clinical agents. The diagnostic agent went through 4 versions.

Problem

Clinical intake needs low-latency, stateful sessions that stay grounded in structured clinical data.

What I built

A WebSocket diagnostic agent on the OpenAI Agents SDK that maps symptoms to clinical Finding IDs through RAG over a custom vector store. A voice intake agent on the Realtime API for paediatric nutrition assessments.

Key decision

Structured outputs, so answers stay in a fixed clinical format that can be stored and used downstream.

• OpenAI Agents SDK
• OpenAI Realtime API
• WebSockets
• RAG
• Vector stores
• AWS Lambda
• DynamoDB
• Cognito

Average API response time fell by about 50%.

Problem

Legacy Java Spring Boot services had to move to Python without getting slower or less correct.

What I built

Redesigned the database schemas and rewrote the core APIs in FastAPI.

Key decision

Response-time goals shaped the schemas from the first draft.

• Python
• FastAPI
• SQL schema design

An installable notes app that works offline and syncs when it can, on a near-zero-cost serverless backend.

Problem

A notes app should open and save instantly with no connection, then sync across devices without losing edits.

What I built

IndexedDB is the source of truth. A background sync engine reconciles with last-write-wins, tombstones and a clock-skew-safe pull cursor. The backend runs on FastAPI, Lambda, DynamoDB and S3.

Key decision

IndexedDB as the source of truth, so notes work offline and the server only keeps devices in step. The API is covered by 23 pytest tests.

• React
• TypeScript
• IndexedDB
• FastAPI
• AWS Lambda
• DynamoDB
• S3
• AWS SAM
• SSM
• pytest