Ask Danny

Describe the job. Get the steps, the parts, the cost, and where to find each part.

ask-danny-ai.com · Live, free beta

---

Trades lose hours and margin on the second trip to the supply store — the fitting that didn't fit, the tool left back at the shop, the size that was wrong.

Ask Danny takes a job described the way someone actually says it:

"My toilet keeps running"

and returns the steps to fix it plus one list: materials with quantities and specs, the tools needed, a rough price, and where each part is at Home Depot, Lowe's, or AutoZone.

The two problems this has to solve

Models invent products. A general model will happily produce a plausible part number for something that does not exist. So the server retrieves from a curated inventory (products.json, 249 products across 10 trade buckets) and injects the matches into its own prompt. Anything not in the inventory is included but marked low confidence.

Aisle numbers get the same treatment. A number is shown only when someone has checked it in that store; the model is told never to give one, and any that slips through is dropped. Every other part shows its department and section ("Plumbing — toilet repair"), and in four mapped Los Angeles Home Depots, which row that department is in and what it sits next to. No store has been checked yet, so today every part shows a department. The goal is 75–80% of shown aisle numbers correct at one store, and that ratio is the number worth improving before anything else.

Models sell the expensive part. Asked why an AC runs but does not cool, a model will reach for refrigerant. The usual answer is a $20 capacitor. The system prompt carries per-trade check-order guidance — capacitor before refrigerant, flapper before a new toilet, coils before a compressor, battery before a starter, jam wrench before a new disposal — because selling someone the costly part when a cheap one usually fixes it is the worst thing this tool can do.

There is also a hard stop above all of it: for a gas smell, a suspected leak, carbon monoxide, sparking or burning wiring, or standing water near live electricity, it returns no materials at all and says to leave and call the utility or the fire department. A short list of nothing plus the right instruction is the correct answer.

How an answer gets made

Several models answer independently, then one synthesises. This is not majority voting — a judge reads the drafts and produces a single list, weighing safety, technical correctness and compatibility.

Browser (index.html)
    │
    └── POST /api/council
           │
           ▼
    netlify/functions/ai-council.js
           │
           ├── retrieval over products.json        (server-side, not client-supplied)
           ├── system prompt built server-side     (never accepted from the request)
           ├── persistent rate limit               (Netlify Blobs — not enforcing yet, see Security)
           │
           ├──► Gemini  ─┐
           ├──► Groq    ─┼──► drafts
           │             │
           └──► judge ───┘  synthesises one list
                    │
                    └── aisle numbers only from checked store data

If the council cannot be reached the client falls back to /api/ai, a single-model path that carries the same server-owned prompt and the same safety rules.

A circuit breaker opens after three consecutive failures from a provider and retries with a single trial request after a cooldown, so one sick provider does not slow every request.

Features

Voice input	Describe the job hands-free (Web Speech API)
Repair steps	3 to 8 steps, safety first; none for hazards, only safe checks for pro work
Photo and video input	Show the problem instead of describing it; a video becomes four frames on the device
Store layouts	Where each department is in four mapped LA Home Depots, after asking for location once
Multi-store compare	Home Depot vs Lowe's, side by side
Tool checklist	What to bring, per trade
Offline mode	Last result cached for the truck
Quantity calculator	Coverage, runs and counts
Job history	Firestore when signed in, localStorage as a guest
Usage ring	How many free lists are left today

Trades: Plumbing · HVAC · Carpentry · Auto · Appliance

Plans: free is 5 lists a day with up to 4 drafters, all on free-tier providers. Pro (40/hour, paid models) is built but not sold — verifyProToken() returns false until real accounts exist, and the tier is never taken from the request.

Evaluation

Architecture is easy to add and hard to justify. The thing that actually tells you whether this works is a set of real jobs and someone qualified marking the answers.

Path	What it is
eval/jobs.json	58 jobs phrased as real people speak, 13 of them traps
eval/run.js	Runs them against the live API, writes a CSV with columns for a tradesperson to grade
eval/local-check.js	200 checks against mocked providers — every deploy runs them and stops on a failure
eval/run.js --compare claude	Runs every job through the free council and through Claude (Opus 5) with the identical prompt, and writes a blind A/B sheet for a grader plus a separate answer key. Answers whether the paid model is worth it before anyone pays for it.

The traps are the interesting part: a gas smell must refuse rather than sell parts, an AC that won't cool must reach for the capacitor and not refrigerant, a humming disposal needs a jam wrench and not a new unit, a load-bearing wall needs an engineer, and a prompt injection must produce a normal list and leak nothing.

local-check.js runs both handlers in-process against scripted providers and has caught two ReferenceErrors that node --check could not see. Run it before every deploy.

Security

• No API key reaches the browser. Keys are resolved server-side from an explicit allowlist of environment variable names, and the value must carry the right prefix — so an unrelated secret cannot be picked up and sent to a model provider.
• The system prompt is owned by the server. Both endpoints previously accepted body.system and forwarded it, which let anyone replace the instructions wholesale — safety rules included. It is now read and discarded.
• CORS is an allowlist, on both endpoints. /api/ai was *, which let any website spend our provider quota from a browser.
• Verified aisles come from our own data. The client used to post verifiedAisles and the server trusted them, so a crafted request could assert any aisle it liked.
• Per-IP rate limiting is stored in Netlify Blobs. Until September 2026 it silently enforced nothing. These are Lambda-compatibility functions, which get no Blobs environment unless the handler installs it from the request, and the library's own connectLambda drops the URL that strong-consistency reads need. Each handler now installs it itself, and local-check.js tests that.
• The update list is confirmed opt-in. A sign-up stays pending until its owner clicks the link in a confirmation email, and the link opens a page with a button rather than confirming on open, because mail scanners open every link.
• Subscriber addresses are encrypted at rest with AES-256-GCM when LIST_KEY is set, so a copy of the store, or an export taken with a stolen admin token, is ciphertext. The blob key stays a hash of the address, so lookups never decrypt. Set it with netlify env:set LIST_KEY "$(openssl rand -base64 32)".
• Claude security review reads every pull request's changes and comments on security problems (.github/workflows/claude-security-review.yml). It needs a CLAUDE_API_KEY repository secret and skips itself without one; fork and Dependabot PRs never get the key.
• CI security checks run on every push and weekly: gitleaks for secrets in the whole history, OSV-Scanner and npm audit for vulnerable dependencies, CodeQL for the code itself, plus the same pre-deploy checks. Dependabot opens update PRs. See .github/workflows/security.yml.
• Input sanitisation, and a token cap enforced against the plan rather than the request.
• Firestore rules restrict every document to its owner and cap field sizes.
• netlify.toml sets HSTS, X-Frame-Options: DENY, , a strict referrer policy, and a permissions policy denying camera, payment, USB and motion sensors.

CSP: script-src allows no 'unsafe-inline'. Every inline onclick was replaced with delegated listeners, and the three remaining inline scripts are allowed by SHA-256 hash. eval/csp-hashes.js regenerates the hashes and local-check.js fails if they drift, because a stale hash blocks all JavaScript. style-src still allows 'unsafe-inline' for the inline CSS, which is lower risk since styles cannot execute.

To report a vulnerability, see SECURITY.md.

Privacy and disclosure

Analytics load only after the visitor accepts — Consent Mode defaults every category to denied before any tag can run, and declining is the same size and weight as accepting. Every user acknowledges that AI can be wrong and that they are 18 or older before using the app, and each result carries an AI-generated label on the result itself, so a printed or screenshotted list still says a machine wrote it.

Photos are sent to an AI provider to answer one request and are not stored. No face recognition, no biometric extraction. Job history for guests never leaves the device.

See privacy.html, terms.html and cookies.html.

Running locally

python3 -m http.server 4173

Then open http://localhost:4173/index.html.

On localhost the page talks to a local Ollama model instead of the deployed functions (USE_LOCAL is set from the hostname), so the AI features need Ollama running.

Before any deploy:

node eval/local-check.js

Layout

Path	What it is
index.html	The entire client app
products.json	Curated per-trade inventory used for retrieval
netlify/functions/ai-council.js	The council: retrieval, prompt, drafters, judge, rate limiting
netlify/functions/ai-proxy.js	Single-model fallback, same server-owned prompt
eval/	Job set, live harness, and pre-deploy checks
netlify.toml	Security headers, CSP, redirects
firestore.rules	Owner-only Firestore access rules
privacy.html terms.html cookies.html	Disclosure, DMCA agent, AI policy

Status

Live and free while in beta. Accounts are not wired up yet — Firebase config is still a placeholder, so everyone is a guest and history stays on the device.

Material lists are generated by AI and are not checked by a licensed tradesperson. Verify quantities, specs and store locations before buying, and hire a professional for gas, electrical or structural work.

Third-party assets

Asset	Source	Licence
DM Sans	Google Fonts	SIL Open Font License 1.1
Interface icons	Feather Icons, inlined as SVG	MIT
og-image.png	Generated by this repo's own script	Ours
Product data in products.json	Compiled by hand for this project	Ours

Store names (Home Depot, Lowe's, AutoZone) and their colours are used to identify where an item is sold. No affiliation or endorsement is implied.

License

Source-available, not open source. See LICENSE.

© 2026 Matthew John Sanchez. All rights reserved.