The model suggests.Code checks.A person decides.

Problem. Accessibility audits check rules, not whether a person can actually finish buying or booking.

• A client never sees a finding until a trainer confirms it
• The hosted worker refuses private and loopback addresses through an egress proxy
• The site is its own release gate: the platform's personas test it before deploy

Stack TypeScript, pnpm + turbo, Fastify 5, Better Auth, BullMQ, Playwright, Prisma, Next 14, SwiftUI

Problem. Coding the mechanism, place, activity and status of injuries consistently across thousands of notes.

• The model never emits a code: evidence spans feed candidates, and a rules engine produces the code
• A simple word-list baseline fails the gate on unseen notes, so it cannot ship on its own
• Policy questions stay marked unapproved until a government reviewer decides

Stack Python 3.14, FastAPI, pydantic, SQLite (Postgres-portable), Qwen3-4B (local), vanilla JS, Docker / mTLS, k3s

Problem. Finding which policies conflict with, already comply with, or cite a changed policy is slow and error-prone by hand.

• The chat model never decides; a code pipeline builds the answer
• The only network exit is one gateway container; everything else sits on internal networks
• Approval comes only from the person's typed 'decide' message, so the model cannot approve itself

Stack Docker Compose, vLLM, IBM Granite 4.2 30B, FastAPI, Postgres + pgvector, Open WebUI, TEI embeddings, LangGraph, Hermes Agent, nginx

Source code

Problem. Sensitive sites cannot let raw prompts, code or data leave, but their engineers still need AI help.

• Redaction uses stable placeholders so answers can be restored locally
• Bundles are signed with ssh-keygen and installed by atomic symlink swap
• Managed settings lock telemetry, updates and marketplace installs off

Stack Python 3 (standard library only), MCP, AWS Bedrock GovCloud, ssh-keygen signing, VS Code extension

Problem. Students and advisors struggle to navigate catalog rules, prerequisites and scattered web pages.

• Keyword search first, vectors optional, because it already ranked the right catalog page first
• Answers come from the prerequisite graph or templates; the local model only phrases them
• No hosted AI services; everything runs on the local machine

Stack Python 3.14, SQLite FTS5, sqlite-vec, bge-small embeddings, Qwen3-4B via mlx-lm, FastAPI, SSE

Open it

Problem. People in an emergency need fast, structured guidance with clear limits on what the AI can claim.

• Reliability, error transparency and accessibility rank above looks
• Every AI output carries the disclaimer

Stack React 18, TypeScript, Vite, Supabase, Perplexity API, Gemini Vision, jsPDF

Problem. Small businesses lack time to plan content and judge ad performance.

• Paid ads never go live without a human: approval is a separate function
• Same edge-function shape as Little Consult

Stack Supabase edge functions (Deno), Vite, React, GitHub Actions

Problem. Existing agent frameworks are too large to audit and weak on security.

• Built on a small framework instead of forking a roughly 928,000-line one
• Tool failures return an error string and never end the run
• Chat watching uses Claude Code's HTTP hooks on 127.0.0.1 only

Stack Python 3.14, Pydantic AI 2.55, MCP, SQLite, git worktrees, Claude Code

Problem. Notes, plans and bank statements live in separate places and never connect.

• Backend is locked; the mobile app never changes the schema
• Every user table filters by the signed-in user's ID

Stack TanStack Start, Supabase, Capacitor, Expo, Plaid

Problem. Each school needs its own searchable knowledge base without rebuilding the product.

• Same codebase copied per customer; only a config file changes
• Crawler hosts allowed by domain suffix, minus login and learning-system prefixes
• Each page is committed as it is fetched, so a partial crawl is still usable

Stack Python 3.14, SQLite, httpx, BeautifulSoup, Windows desktop

Problem. Founders need ongoing guidance, not a one-off report.

• Reuses Little Consult's edge functions and health-check funnel
• Documentation kept in sync with marketing claims

Stack Vite, React, shadcn, Supabase

Problem. Finding contactable prospects at scale without guessing emails.

• Emails are never guessed; each row stores its source URL
• Personal-data flags are added per row

Stack Python, openpyxl, Subagents

Problem. Phone apps needed a private assistant without sending knowledge to a hosted AI service.

• Retrieval is always injected first because the 14B model's tool choice is unreliable
• Each app's bearer key reaches only its own knowledge base
• Replies stream as server-sent events: token, sources, proposal

Stack Ollama, FastAPI, SQLite FTS5, numpy, Caddy, Cloudflare Tunnel

Problem. Early ideas get little diverse, critical feedback before money is spent.

• Red team is scored separately so attackers do not drag down the opinion mean
• A region grid forces a third of each batch to be traditional or right-of-centre, because ungrided output drifted toward Western liberal professionals

Stack Python 3.14, MCP 2.2, Markdown vault

Problem. When a sector policy changes, every entity below it needs checking, and each reader may only see their own scope.

• Database uses row-level security with an app role that cannot bypass it
• Each signed link token gets its own key and purpose after a security review
• A rebuild branch removes the large agent framework in favor of a small LangGraph core

Stack FastAPI, LangGraph, Postgres 16 + pgvector (RLS), vLLM, TEI embeddings, Open WebUI, nginx, Docker

Source code

Problem. People need quick answers grounded in official text, with human routing for anything binding.

• Knowledge base is Markdown; guardrails sit in the system prompt (no invented deadlines, crisis questions to hotlines)
• Local similarity matching reduces API dependence over time

Stack Next.js 16, Perplexity API, BM25, Node.js / Express, Anthropic SDK, TF-IDF