Projects built to make complex systems usable.

3-node linear pipeline (retrieve_semantic -> retrieve_graph -> analyze) that does semantic + graph retrieval then a single LLM call to produce a cited analysis (general/SWOT/competitive/trend). Self-labeled 'Research agent' with an 'A-RAG adaptive' docstring, but the implementation is a fixed linear sequence, not dynamic tool-choice.

custom

3-node pipeline (gather_findings -> extract_brief -> generate_outputs) that runs an LLM map-reduce extraction into a StrategyBrief, then fans out to format generators (pptx/docx/email/notes/outline).

custom

Top-level orchestrator (Opus): plans via research-planner, delegates to specialist subagents, tracks progress, enforces Pyramid Principle output and artifact verification

claude-agent-sdk

Decomposes research questions into structured task plans before data gathering

claude-agent-sdk

Extracts and verifies data from PDFs, websites, spreadsheets, and presentations

claude-agent-sdk

Analyzes SEC filings, earnings reports, and financial statements; runs Python calculations

claude-agent-sdk

Synthesizes multi-source extracted data into MECE/Pyramid-structured narratives with citations

claude-agent-sdk

Produces executive-ready SCQA briefs from synthesis output

claude-agent-sdk

Designs process architectures and flowchart visualizations as PowerPoint decks

claude-agent-sdk

Designs visual layout/structure for deliverables applying Calm Precision principles

claude-agent-sdk

Refines weak/ambiguous research prompts for accuracy and specificity

claude-agent-sdk

Monitors execution for scope/quality/source/goal drift against the research plan

claude-agent-sdk

Verifies synthesized claims against cited sources; flags unsupported claims and stale data

claude-agent-sdk

Scores report quality (completeness, accuracy, structure, depth, actionability) against a 4.0/5.0 threshold

claude-agent-sdk

Post-run debrief that diagnoses execution quality and records lessons/memory updates for future runs

claude-agent-sdk

Lead-agent session driver built on claude-agent-sdk's query(); TeammateManager (2039 lines) spawns, messages, and manages the lifecycle (active/idle/resume) of sub-agent 'teammates' the lead agent creates at runtime.

claude-agent-sdk

MCP server + service letting a running agent session create/spawn/inspect/prompt sibling or child agent sessions across providers (Claude, Codex) and worktrees (tools: create_session, spawn_session, get_session_status/result, send_prompt, respond_to_prompt, list_spawned_sessions).

mcp

Main production research agent — DeepAgents single-agent flow with built-in write_todos planning, filesystem tools, and task-tool subagent delegation (planner/researcher/analyzer/writer/fact_checker/reviewer/reflector/worker role prompts)

langgraph

Lead agent decomposes task into worker subtasks, workers execute independently and register artifacts, lead synthesizes from artifact summaries only (telephone-game mitigation, one-level-only enforced)

custom

Evaluator-optimizer loop — writer agent generates content, critic agent scores/verdicts, writer revises until score threshold or max iterations

custom

Reflexion-style loop — same-model agent generates an answer, then self-critiques against a rubric and revises (bounded iterations, stagnation guard)

custom

Runs N sampled agent turns at varied temperature concurrently, then a synthesizer agent merges them into one answer (ensemble voting / hallucination catching)

custom

Classifier turn labels the query, then dispatches to one specialist agent turn scoped to that label

custom

Linear staged agent pipeline (gather -> analyze -> write), each stage consuming the prior stage's output, optionally pinned to different models per stage

custom

Multi-node orchestrator for schedule optimization, priority triage, time-block planning, decision logging, and follow-up tracking with role-based task assignment.

custom

Document analysis and knowledge management agent with embeddings, retrieval, and local LLM summarization for meetings, documents, and research materials.

custom

Planning agent that constructs structured outlines, generates plan sections, and builds implementation roadmaps using local LLMs.

custom

Meeting analysis agent that extracts actions, decisions, and notes from meeting transcripts with semantic classification.

custom

Orchestrates LLM requests between on-device (Apple Foundation Models) and cloud (Groq) providers with routing, caching, and state management for AI features

custom

Orchestrates plan critique generation by combining deterministic PlanCoachEngine logic with LLM inference and safety validation

custom

Orchestrates weekly review generation via AIInsightService with weekly caching, permission checks, and prompt construction

custom

Routes task decomposition between Apple on-device Foundation Models and cloud providers for parsing raw text into structured tasks

custom

Multi-step intelligent-search pipeline — classifyIntent → planRetrieval → fan-out retrieval (vector+KG) → rerank → quality-guard (reformulate-and-retry on low relevance) → synthesize → numeric verification → normalize; feature-flagged, falls back to legacy search handler.

custom

Conversational multi-step router: detects specialty/intent, generates and adapts intake surveys, classifies via Stage 0, then dispatches to the Decision Engine or Recommendation Engine and streams the Groq completion back to the user

custom

Deterministic multi-stage decision pipeline (weights/AHP -> constraints -> weighting -> outranking -> ranking -> feasibility -> scaffold), with LLM calls at Stages 1, 5, 6 via Groq

custom

Chains pain-path classification -> library retrieval -> candidate task generation -> 9-criteria scoring -> promotion (or the newer runWorkflowV2 decompose/score/lynchpin/horizon pipeline) to produce an AiTaskRecommendation

custom

Multi-strategy content scraping orchestrator supporting RSS discovery, sitemap parsing, HTML scraping with fallback logic.

custom

Multi-step orchestrator that runs guided-conversation intake for a product spec: routes to a method (jtbd/qfd/pugh/agent) via a Haiku sub-call, scores unknowns 0-15 to decide ask-vs-infer, generates the next question or safe-default assumptions, then finalizes into a rendered spec draft.

custom