Python backends, data pipelines and ML that hold up in production.

Finding faults in residential solar systems without a site visit

3S wanted to know when a customer's PV system produces less than it should, and why, using only the power data the system already sends. We built the analytics platform that does this every night.

• Fault detection and diagnosis engine. Physical PV modelling with pvlib and horizon profiles, a clear-sky day classifier, trend analysis and stuck-value checks. Each loss is checked against weather, snow and hail data before it is reported.
• Performance Loss Rate (PLR). Multi-year, year-on-year degradation analysis, recalculated every month.
• Public REST API. Endpoints for performance, trends, outliers, hail history and daily hail risk, with scoped API keys and per-scope rate limits.
• Data pipelines. PV data arrives through AWS SQS every 30 minutes. Weather, irradiance, snow and precipitation come from several providers. Celery runs the nightly jobs.
• Regression baselines. Reference systems with pinned results, so a library upgrade or algorithm change never moves the numbers by accident.

• Python
• FastAPI
• SQLAlchemy
• PostgreSQL
• Celery
• Redis
• pvlib
• pandas
• scikit-learn
• AWS SQS · RDS · ECS
• Docker

Running PV analytics as a scalable service on Kubernetes

Backend and data engineering for Solar Manager's energy management platform, which connects home solar, batteries, heat pumps and EV chargers.

• Re-architecture. Moved the analytics service from one container running six processes to separate Kubernetes workloads: API, batch workers, data ingest, scheduler and a migration job that runs before each deploy.
• Data continuity. A site keeps its full history when its hardware is replaced or the system moves, and earlier fault data is merged in.
• Safe changes. A local sandbox with fake upstream APIs for scenario tests, versioned releases and metrics in Prometheus.

This work is under NDA. Happy to talk about the approach in a call.

• Python
• FastAPI
• Celery
• Valkey
• PostgreSQL
• AWS EKS
• Kubernetes
• Terraform
• Prometheus
• GitLab CI

Scoring spoken English tests with machine learning

For a large language testing provider, human raters score recorded speaking tests. The goal was automated scores that agree with raters at least as well as raters agree with each other, plus useful written feedback for every learner.

• Proficiency scoring model. DeBERTa embeddings for each answer plus 24 prosody features from the audio, combined with attention pooling across the 8 to 14 answers in a test.
• Rehearsed answer detection. A classifier trained on about 80,000 responses that flags memorised answers, using text together with 36 audio features.
• Diagnostic reports with an LLM. An event-driven pipeline on AWS that writes a feedback report after each test session. Duplicate triggers are safe, failed runs are tracked and retried.
• Prompt control. Prompts are versioned by hash and a CI check blocks a merge if a prompt changes without review.

• PyTorch
• Hugging Face
• DeBERTa
• XGBoost
• FastAPI
• AWS Lambda
• Step Functions
• SQS
• Amazon Bedrock (Claude)
• Terraform

Reading contracts at scale with OCR and LLMs

Large volumes of contracts and public records had to be collected, read and classified without manual work.

• PDF ingestion pipeline. OCR, page splitting, validation and structured extraction.
• LLM classifiers. Contract type detection, validation and addendum detection with GPT and Claude.
• Async scrapers. Playwright jobs with strict validation and error handling.
• Reliability. Sentry alert rules, error categories, dashboards and trace IDs across services.

• Python
• FastAPI
• Celery
• Playwright
• OCR
• OpenAI
• Anthropic
• Google Drive / Sheets API
• Sentry