Projects

Lenny Compass

Lenny Compass helps readers answer product questions using Lenny’s articles and podcasts, with original passages available to check each claim. It brings questions, search, source reading and citation checks into one local workspace, reducing repeated searches across scattered material. Quick mode retrieves evidence for focused questions; deep mode lets an agent search and select evidence over several rounds. A verification step checks claims against their citations and repairs or omits unsupported statements. Readers can open the surrounding source text, distinguish corpus findings from model suggestions, save conversations and ask follow-up questions, keeping a research trail they can revisit.

Tweets2Insights

Tweets2Insights combines statistical patterns in posts with external facts to produce public-opinion reports whose conclusions can be traced to evidence. A data agent analyzes labeled records with quantitative tools, a search agent gathers external facts, and a coordinator organizes further evidence gathering, cross-examination and conflict detection. This workflow addresses the narrow evidence base of a single-agent analysis. Core report claims are linked to sources and revisited by a verification agent. A visual interface exposes task steps, tool inputs and outputs, and scheduling paths, so users can investigate how a conclusion was reached or locate a problem in the analysis.

Novel Evaluation

This local fiction-evaluation system produces reports with explicit criteria and inspectable judgments, helping readers understand scores and reuse a rubric across works. It screens the input, selects an analysis mode, identifies the work’s type and scores eight dimensions, then adds type-specific assessments and checks consistency before compiling the report. Users can review individual judgments, adjust prompts and criteria, and compare results through saved history. Task creation, batch evaluation, output-schema validation and regression tools support repeated checks of the evaluation process. The report exposes the reasoning behind the overall score so that its criteria and tradeoffs can be discussed and revised.

LaterOn

LaterOn turns scattered research pages into a project library that readers can return to, reducing tab clutter and repeated searches. I extended its collection and reading workflow with AI-assisted organization: pages are saved first, then read and assigned using project names and descriptions. Extraction can run locally in the browser or through a remote service; classification supports Chat models and decision-model interfaces such as Jev and OpenAI Decisions. Pages without a clear match remain in the inbox for manual review. Full-screen and sidebar views preserve reading status and scroll position. It addresses the same organizing problem as AutoTabGroups: LaterOn supports a lasting, separate library, while AutoTabGroups organizes native tab groups in the current browsing window.

AutoTabGroups

AutoTabGroups organizes scattered pages in the current window into native browser tab groups, reducing manual sorting. It uses page titles, descriptions and optional body text alongside group names and descriptions to choose a destination. Existing groups take priority, with new groups created when requested. Already grouped pages keep their placement, and uncertain matches remain untouched. Users can inspect reasons and candidate scores, correct assignments, request a new judgment or undo the latest grouping change. Page reading and classification are configured separately, with Chat and decision-model options. Like LaterOn, it explores AI-assisted webpage organization. AutoTabGroups works directly with Chrome tab groups for active browsing tasks; LaterOn uses a separate project library for saved material and later reading.