Company · Hoxton, London · 5 days a week in person
Tracelight
Tracelight is Cursor for serious financial modelling in Excel: an AI add-in that builds, reviews and stress-tests the models companies make their biggest decisions on. Long term the team is building an AI operating system for the whole strategic decision lifecycle: the work that today runs across Excel, PowerPoint, email and meetings.
- Headquarters
- Hoxton, London · 5 days a week in person Five days a week in person
- Product
- Tracelight Platform, Excel add-in, PPT & Word add-ins, Spreadsheet Apps, and Model Review, across commercial and financial due diligence, M&A, valuations and modelling, private equity and corporate development.
- Team
- A small, very strong engineering team described as ex Jane Street, Bloomberg, Cambridge and UCL. TypeScript for most things, Rust where it matters. SOC 2 Type II, GDPR and ISO 27001 on the security page.
Open review
One role, reviewed in full
Each role gets its own page, run as a model review against the posting.
Why this one
The argument about
numbers you can trust
Tracelight's thesis is that financial modelling is the worst-tooled form of programming, and that a model nobody has independently reviewed is a decision made on faith. That is a claim about evidence, not about spreadsheets.
Wiregent arrives at the same claim from the settlement side: a double-entry, hash-chained ledger where every amount is integer cents and every balance is recomputed from the legs on read, so the number is never something you have to trust someone to have updated. Beside it sits an eleven-category evaluation taxonomy whose most important design decision was to report an unevaluable rule as unproven rather than zero.
The posting's second bullet asks for someone who improves the harness, the tools and the context. Here that is a PreToolUse hook intercepting grep and handing the agent a scoped knowledge subgraph, eighty-two skills, five subagents with strict file ownership, and an MCP server written so an agent drives the same library a human does.
See for yourself
Start with the review
Findings graded by severity, scenarios flexed, and the ones that fail named as loudly as the ones that clear.