I build agentic AI workflows for asset-management teams. Process mapping, requirements, tooling, regression-tested implementation, and adoption support. My focus is the operational layer between the user and the data, and the long tail of admin work where practical AI actually pays off.
Agentic AI that reads private-equity quarterly reports and runs a nine-item LP diligence checklist
A multi-step agent that performs LP-side diligence on private-equity fund quarterly reports. Given a quarterly report, it runs a nine-item checklist (NAV drivers, capital called, distributions, IRR/TVPI/DPI, top holdings movement, unfunded commitment, fees, key-person events, valuation methodology) and returns citation-backed answers with explicit confidence tags. Built on Anthropic Claude with prompt caching, sqlite-vec retrieval, an MCP server interface, and a 20-question evaluation harness. The differentiator is the refusal mode: when the report doesn't disclose a figure or the data is redacted, the agent returns 'data not available' rather than fabricating, and refusal correctness is scored as a first-class metric alongside faithfulness. Live demo at lp-diligence.krawczun.com.
An MCP server for querying SEC EDGAR filings, built for AI agents and humans
A Model Context Protocol server backed by a RAG pipeline over SEC EDGAR 10-Ks. Ingests filings directly from EDGAR, splits them into section-aware chunks, embeds locally with sentence-transformers, retrieves via sqlite-vec, and synthesizes grounded answers with inline citations back to the filing. Pluggable across embedding providers (local or OpenAI) and LLM providers (Claude CLI, Anthropic API, OpenAI API). Answers questions like "What did the filing say about AI risks?" against Apple's most recent 10-K end-to-end, with real measured evaluation metrics published in the repo's README. Full-run aggregate scores include faithfulness 0.998, refusal_correctness 0.886, and context_recall 0.443 across a 35-question hand-curated golden set.
Hyperscaler → Supplier Attribution
A working prototype of an AI-enabled research workflow for asset-management analysts. Pulls $400B+ of annual hyperscaler capex directly from SEC EDGAR filings, runs Claude-CLI structured extraction over earnings transcripts, and routes the dollars through to specific supplier exposure via a transparent attribution model. Every figure is source-linked back to its filing, auditable end-to-end. The “Ask the Analyst” chatbot is grounded in the curated dataset, demonstrating a Human + AI design where the model assists rather than replaces analyst judgment.
Daily RNS share-transaction monitoring with a self-improving extraction pipeline
Runs daily against the London Stock Exchange RNS feed, pulling every UK Investment Trust share-transaction announcement of the day: buybacks, issuances, tender offers. Claude CLI is the primary extractor. A deterministic regex layer is the fast path for boilerplate formats, and patterns Claude and regex agree on for three consecutive runs get promoted into the regex library, so the system gets faster and cheaper over time without code changes. Built as a regulatory-grade instrument: output fed daily TR1 disclosure decisions under FCA 48-hour thresholds, replacing ~2 hours per day of manual review.
Find Active ETFs that match a stock or mutual fund
Enter any US-listed stock, ETF, or popular mutual fund and rank the Active ETFs most similar to it by underlying exposure. Builds per-fund profiles from SEC N-PORT-P filings, scores across holdings overlap, sector and market-cap tilts, style factors, and 3-year weekly return correlation. The corpus is rebuilt offline; the live tool is fully static and runs the similarity math in the browser.
Workflow-aware document review for regulated teams
A document review and approval platform for small regulated teams, built around the workflow itself, not the document. Every artifact moves through a defined review sequence with role-based approvals, parallel stages, change-request paths, and a complete audit trail. The wedge over generic e-signature or CLM tools: workflow-aware compliance evidence. AI review hooks plug into the stage state, so model-assisted checks become a controllable step rather than an opaque pre-process.