30+ production AI systems shipped.

Agent platform

Drives 100+ agents from one chat window.

• MCP
• TOOL CALLING

Production agents

Shipped as MCP microservices.

• MCP

Agent library

70+ tool-using agents across search, travel, finance, media and productivity, each its own MCP server.

• MCP
• NODE.JS

Financial-advisor assistant

With its own tools for portfolios, trade status, market outlook and money movement.

• TOOL CALLING
• FASTAPI

Health-guidance assistant

On WhatsApp, grounded in vetted sources, handing over to a person when it should.

• WHATSAPP
• RAG
• RERANKING

Team check-in agent

Collects each team member’s status by chat and feeds the owner’s dashboard.

• SCHEDULED AGENT
• SUPABASE

Multi-agent build pipeline

One prompt becomes requirements, a sprint plan, a repository and features built in parallel by separate agent roles, merged automatically.

• MULTI-AGENT
• FASTAPI
• GIT

Career-intelligence platform

AI-disruption risk scored from a CV or profile, with a career knowledge graph, job matching and next-role recommendations.

• NEO4J
• QDRANT
• PREFECT

Multi-agent writing platform

Writer, editor, proofreader and researcher agents, with each job routed to a suitable model provider.

• MULTI-AGENT
• MODEL ROUTING

AI video platform for ads

Each job goes to a suitable video model, renders on GPUs in a queue, and opens in a browser editor for changes.

• DIFFUSION
• GPU QUEUES

Virtual try-on for fashion

A garment image and a person image go in, and the garment comes out on the person, on a diffusion pipeline.

• VTON
• DIFFUSION

Image suite

Background removal, upscaling, face re-ageing, pose estimation and garment segmentation behind one interface.

• PYTORCH
• OPENCV

Podcast platform

Turns articles and data into scripted, narrated episodes with a ranked social feed.

• TTS
• FASTAPI
• S3

70B open-weight model

Fine-tuned with SFT and DPO.

• SFT + DPO
• VLLM

GRPO for function calling

Post-training with our own reward function, evaluated on a public benchmark.

• GRPO
• TRL
• VLLM

Reasoning models

Post-trained with GRPO, SFT and a process reward model, scored on MATH-500 on 8-GPU runs.

• GRPO
• PRM
• TORCHRUN

Swiss-tournament Elo leaderboard

Ranks language models on maths and penalises hallucination.

• ELO
• VLLM
• SGLANG

Pretraining 3B-class models

Pretraining and continued pretraining on multi-node GPU clusters.

• DISTRIBUTED TRAINING
• MULTI-NODE GPU

Self-play preference trainer

Writes its own contrast data and tunes with DPO on LoRA.

• DPO
• LORA

Private LLM and SLM deployments

Open-weight models, with quantised serving.

• OPEN WEIGHTS
• QUANTISATION
• VLLM

Speech-to-speech assistant

Starts speaking before the model finishes.

• WHISPER
• KOKORO TTS
• TRITON

NLP research and evaluation

Peer-reviewed research at a top venue, and a published multi-agent framework for evaluating language and vision-language models.

• RESEARCH
• EVALUATION

Knowledge-graph question answering

Graphs built from facts a language model extracts, with graph embeddings to help it reason.

• KNOWLEDGE GRAPH
• EMBEDDINGS

Document AI

OCR, tables linked across Word and Excel, and contract and compliance checks.

• PGVECTOR
• NEO4J
• MONGODB
• REDIS

Layout-aware insurance pipeline

Insurance policies read into section-mapped knowledge, with no LLM needed for extraction.

• CELERY
• POSTGRESQL
• S3

Financial report checks

Grammar, currency, arithmetic, units and headings, checked before a report goes out.

• ANOMALY CHECKS

Commercial lease analysis

Clause by clause, with redlines exported to Word.

• CLAUSE ANALYSIS
• DOCX

Scanned building specifications

Turned into structured tables, with search over the whole set.

• PADDLEOCR
• QDRANT

House-style formatting engine

Learns a house style from reference documents and fixes every Word file to match.

• DOCX
• STYLE PROFILES

XBRL matching and workbook comparison

Annual-report pages matched to their tagged XBRL filing, and workbook versions compared cell by cell without false alarms from inserted rows.

• XBRL
• TF-IDF
• LCS

LLM evaluation and leaderboard

With LLM-as-judge and cost tracking.

• LLM-AS-JUDGE

Perplexity-style search

Query planning, retrieval, reranking and cited synthesis.

• RERANKING

Credit-document extraction

Read three ways, by a vision model, in JSON mode and against a schema, and reconciled into one record.

• VISION MODEL
• PYDANTIC

Clinical transcription structuring

Medical transcriptions turned into structured clinical notes through an API.

• FASTAPI
• LLM

Mobile AI app

With 30k users.

• MOBILE AI

macOS command centre

A native app that supervises many coding-agent sessions at once, with a risk-ranked approval queue.

• SWIFTUI
• NATIVE MACOS

AI study workspace

Notes, flashcards with spaced repetition, quizzes, multiplayer quiz battles, mind maps and podcast recaps from a student’s own PDFs, recordings and links.

• NEXT.JS
• OPENAPI
• AWS

Nutrition app

For iOS and Android: meal plans, food-photo scanning, health-report reading and an AI coach.

• EXPO
• FASTIFY
• POSTGRESQL

Client-reporting platform

Six levels of drill-down, goal simulation and narrated video summaries.

• NEXT.JS
• MONTE CARLO
• TTS

Owner and employee dashboards

Live team analytics, focus timers, goals and reviews.

• NEXT.JS
• SUPABASE

Multi-tenant admin consoles

Organisations, credits, feature flags and audit logs for a multi-tenant AI platform.

• MULTI-TENANT
• GRAPHQL

API-key vault

Per-key usage logs and analytics.

• USAGE ANALYTICS

AI-native IDE

Agent chat, retrieval over the codebase, a built-in terminal and OS keychain credentials.

• ELECTRON
• MONACO

Demand-forecasting platform

Hierarchical time series with 150+ engineered features.

• CATBOOST
• LIGHTGBM
• XGBOOST
• AUTOGLUON
• OPTUNA
• SHAP

Identity-document tampering model

A segmentation model trained from scratch that finds tampering in scanned identity documents, pixel by pixel.

• U-NET
• PYTORCH

Microscopy classifier

For materials research, validated with grouped cross-validation and bootstrap confidence intervals.

• CNN
• STATISTICS

Quant research lab

Backtests whole portfolios with full trading costs and records every verdict, failures included.

• BACKTESTING

SCADA display repair engine

For control-room displays migrated between industrial SCADA platforms: code owns every coordinate, and a vision model only identifies and judges.

• VISION MODEL
• GEOMETRY ENGINE

SCADA 3D display converter

Flattened control-room displays rebuilt with native 3D symbol graphics, every live data overlay kept in place.

• VISION MODEL
• SVG

FinOps monitors

Watch AI spend across providers, including dashboards with no API, and alert in Slack and Telegram.

• PLAYWRIGHT
• NODE-CRON

GPU serving on AWS

With containerised AI services.

• DOCKER
• AWS
• BEDROCK
• POLLY
• TRITON

Retrieval and model stack

Infrastructure as code for retrieval and model serving.

• TERRAFORM
• KUBERNETES
• HELM