AI-assisted workflows for
enterprise,
that people actually use.

A working document Q&A agent answered the same 20-question evaluation set in ten configurations: six Claude models from Haiku to Fable, some at several context sizes, with every answer graded against exactly the evidence the model saw. The page prices each configuration from the usage the API actually reported, projects it to a monthly bill at your volume, and recommends the cheapest one that clears a quality bar you set, from the models you choose to consider. Then it lists, by name, the questions the cheaper option gets wrong. The most expensive models tied a mid-tier one at five times the cost, and fixing a grader that saw less than the model changed 27 of 200 verdicts.

Twelve back-office changes scored on four things: how visible each would be to a customer if it failed, how much of the daily transaction volume it touches, how long a silent failure would run before anyone noticed, and whether there is a clean way back. The tool sequences them into release waves with explicit verification gates, then compares that ordering against the one teams actually use, which is quickest wins first. The weights are adjustable because they are judgments rather than constants; the ordering that survives several different weightings is the one worth arguing for.

A multi-step agent that runs a nine-item diligence checklist (NAV drivers, fees, IRR/TVPI/DPI, valuation methodology) against private-equity quarterly reports and returns citation-backed answers with explicit confidence tags. When a report doesn't disclose a figure, the agent returns 'data not available' rather than fabricating.

A planning tool for a national support desk answering four time zones from one location. Each zone contributes the same daily rhythm in its own local time, so the textbook lunch dip fills in and what survives is a five-hour plateau that eight-hour shifts cannot cleanly express. Requirements come from the standard Erlang C method; a roster search runs alongside it to separate the coverage gaps that need hiring from the ones that only need different start times, and prices both against a fully burdened hourly rate. An AI-agent dial shows why routine deflection relieves the plateau without shrinking the team.

Runs daily against a Regulatory News Service feed, parsing every UK Investment Trust share-transaction announcement of the day. Claude CLI is the primary extractor with a deterministic regex fast path, and patterns the two agree on get promoted into the regex library, so the system gets faster and cheaper over time without code changes.

Placet (Latin: "it pleases") is the formal assent by which a governing body approves a matter put before it. A document review and approval platform for small regulated teams, built around the workflow itself, not the document: defined review stages, role-based approvals, parallel sign-offs, and a complete audit trail. AI compliance pre-checks plug into stage state, so model-assisted review is a controllable step rather than an opaque pre-process.

A Model Context Protocol server backed by a RAG pipeline over SEC EDGAR 10-Ks: section-aware chunking, local embeddings, sqlite-vec retrieval, and grounded answers with inline citations back to the filing. Pluggable across embedding and LLM providers, with measured eval metrics published verbatim in the README.

An AI-enabled research workflow for asset-management analysts. Pulls hyperscaler capex directly from SEC EDGAR filings, runs Claude structured extraction over earnings transcripts, and routes the dollars to specific supplier exposure through a transparent attribution model. An “Ask the Analyst” chatbot is grounded in the curated dataset.

Enter any US-listed stock, ETF, or popular mutual fund and rank the Active ETFs most similar to it by underlying exposure: holdings overlap, sector and market-cap tilts, style factors, and 3-year return correlation. Profiles are built offline from SEC N-PORT-P filings; the live tool runs the similarity math in the browser.

Most internal AI tooling dies between the demo and the desk: it needs a key someone has to provision, a server someone has to own, or a login the user never adopts. Shortlist removes all three. Every model call runs through the user's own Claude subscription on her own machine, so there is no API key, no per-token billing, and no server. The output writes straight into a formatted sheet in her Google Drive, and she keeps working in the tool she already used.