Building the tools
agents actually need

A multi-agent system that detects cash and security shortfalls across fixed-income portfolios (T+0 to T+2), then constructs, validates, and books the repo and reverse-repo trades that resolve them. The agents only trade when there's a problem to solve — every proposed trade traces to a specific shortfall on a specific settlement day. A deterministic Python engine does the maths; an adversarial Risk agent runs eight mandatory checks before anything books. Built for the Claude Code Hackathon 2026, drawing directly on 20+ years in collateral and repo. Watch the 3-minute overview

Playwright-based browsing layer purpose-built for agent workflows. TOML configs define extraction patterns; structured outputs go straight into downstream agents. FastAPI MCP server makes Wraith addressable from any agent runtime — Claude, OpenAI, local models. Built because every "just scrape it" task hides a six-month maintenance burden.

Five-dimension evaluation system combining LLM-as-judge with deterministic checks. SQLite drift tracking surfaces when an agent's behaviour has materially changed between deployments. Built because "evals at launch" doesn't survive contact with a model that quietly upgraded on Tuesday. Open-core, with a managed tier from £49/month and advisory at day rate.

18-module autonomous trading system for Pokémon and One Piece TCG markets. Bloomberg terminal-style UI, BYOK LLM router, signal engine, portfolio manager, and human-in-the-loop eBay execution. A serious test bed for the patterns Silsila applies to regulated trading systems — just with cards instead of bonds.

Per-persona RAG pipelines simulate how different stakeholder groups will react to a draft requirement, surfacing contradiction and concern before specifications hit development. Cuts late-cycle rework by finding the objection that would have come up in week eight UAT — in week two discovery.

Production Python service with four specialised agents — Product Owner, Business Analyst, Developer, QA — running an autonomous delivery loop. Linear is the ticket system; Slack is the human interface. Targeted at technical PMs and BAs in regulated industries who need agentic delivery patterns with proper audit trails.

A Claude Code skill with four modes from one entry point: create a CV, improve bullets, review against regional norms, or tailor to a job description. It separates regional archetypes (length, photo, sections, personal-detail rules) from industry overlays (section emphasis, must-have credentials, bullet patterns), so any region composes with any industry at fill time. Markdown source builds to PDF and DOCX via pandoc. Ships with five regions and seven industries in v1. View on GitHub

Gets non-technical people productive with Cowork without writing code or touching settings. Download, unzip, attach the folder, paste one prompt — a bootstrap file then walks the user through setup. Includes a printable Ask-vs-Act card, a plain-English glossary, ten ranked copy-paste first tasks, a connector picker, a when-it-goes-wrong guide, and five ready-made project workspaces. The thesis: the skill to learn isn't perfect prompting, it's steering a confident assistant who occasionally needs correcting. View on GitHub

Drop in a CSV or Excel file and get back a cleaned Tableau extract, a categorised list of data stories, and a polished HTML report — no coding required. A four-stage pipeline profiles, cleanses, narrates, and exports: it detects types, missing values and outliers, fixes them with every decision logged, then has Claude turn the numbers into plain-English stories across eight categories (trends, anomalies, segments, and more), each mapped to a suggested Tableau viz. The audited cleansing trail makes it defensible in regulated settings. View on GitHub