This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

I race 1/10 touring cars in my home country. Almost nothing we run is sold here, so parts come mostly from the US, and every club member has a drawer of spares: a spur gear bought twice, the arms left over from a conversion, an old chassis kit. Meanwhile someone else in the club is asking the WhatsApp group for exactly that part, and the message scrolls away. When nobody has it, the member asks whoever brings parts in, and that person prices the import by hand.

PartsPaddock is an agent that lives in the club's WhatsApp group and does both jobs.

• Swap and sell inside the club (free, P2P). "Tengo un spur 68T de TT-02 nuevo, $10" becomes a listing. "@bot busco un TAM54500" becomes a request. When a request meets a listing, the bot writes to both members in private.
• Import quotes (paid, B2C). When the club has nothing, the bot looks the part up at an online hobby shop and answers with the store price, the weight the store lists, our per-pound fee, and two totals: one if the member orders it and ships it to our forwarding address, one if we order it for them (+12%). Once the member wants to confirm the purchase, if we order it, they need to send a deposit receipt PDF, which the agent forwards to the manager DMs to confirm the deposit; if they order it themselves, they send the store's tracking number. Confirmed import requests then go into the next order batch (e.g. daily batches). The manager gets one summary per batch with links, quantities and what to charge whom.

Who pays: the club member who wants a part brought in. The fee per pound and the 12% for placing the order are the side hustle; the free matching is what keeps the bot in the group. It is small and specific on purpose: one club, one store, one person running imports.

Demo

Code

The project is in challenges/03-week2-side-hustle. MIT licensed.

How I Built It

Three pieces, and a strict line between what the model does and what code does.

The agent: Hermes Agent with Gemma 4. Hermes Agent is an open-source agent harness from Nous Research. It already has a WhatsApp gateway (a Baileys bridge that links a number like WhatsApp Web), skills, cron jobs and per-profile configuration. The bot is a Hermes profile distribution: a persona file, a config, two skills (rc-listings and rc-quote) and a manifest, so the whole agent installs on another Hermes machine with one script. The model is Gemma 4 26B A4B, an open-weight mixture-of-experts model with about 4B active parameters, served by OpenRouter. The bot has a dedicated prepaid SIM, answers in the group only when mentioned, and only allowlisted numbers can talk to it.

The skills: two SKILL.md files, one per business. A Hermes skill is a markdown file with a short description the agent sees in every turn; when a message fits it, the agent opens the skill and follows it. I split the bot into two skills because the free side and the paid side have different rules, and a smaller skill is easier for a mid-size model to follow than one long manual.

• rc-listings (P2P, free). What: recording what members have and need, finding matches, answering "who has X?", and closing a listing when a member says "ya lo vendí". Why: this is the only place where the model's language skills matter: it has to turn "tengo un spur 68T de TT-02 nuevo, lo doy en 10 dólares" or "I'm looking for a rear lower arm for the X4" into one small JSON object. How: the skill gives the JSON shape with an example, says "copy the part number only if the member wrote one, never invent it", and maps each command result to one next step: ok → send reply as written; missing_fields → ask the one question in reply, then run it again; ambiguous → ask which listing. Validation, part-number normalization, matching, the private match messages and the "sold" notice to the other member all happen in rcparts listing add / search / listing close.
• rc-quote (B2C, paid). What: import quotes from the hobby shop, accepting a quote, the deposit receipt, the self-order tracking number, and the maintainer's "confirmar" / "rechazar". Why: every number here is money, so the model must never compute or round anything, and the steps that move money (paid, confirmed) need a person. How: the skill tells the model to pass the member's words to rcparts quote, quote accept, deposit submit --file <the PDF path Hermes gives it> and tracking submit --text "<their message>", and to send reply verbatim. The quote math, the two totals, reading the amount and reference from the PDF, the carrier from the tracking number, and the rule that only the maintainer's number can confirm a payment are code. The model cannot get around that rule, because the sender comes from the WhatsApp session, not from what the model passes. A quote asked in the group is sent by DM, so prices and the forwarding address never sit in the group chat, and "mis pedidos" lists a member's confirmed purchases with their status and tracking number.

Around the skills, SOUL.md holds the persona and the rules that apply everywhere: answer in the member's language, keep group replies to two lines, never post phone numbers in the group.

The code: a small Python CLI, rcparts. The skills tell the model to run rcparts listing add --json '…' or rcparts quote --query TAM54500 and to send back the reply field as written. Every command prints one JSON object with ok, reply and action (the next step, which a mid-size model follows better than a paragraph in a skill). The CLI validates the listing, normalizes part numbers, saves it, finds matches, computes quotes from the store's own JSON, and sends the private messages with hermes send. The order batch is a Hermes cron job that runs a script with no model turn at all.

The data: Postgres on Tiger Cloud. Clubs, members, listings, matches, quotes, deposits and orders live in a Tiger Cloud service. Matching is an exact part-number match first, then Postgres full-text search over the glossary concepts (with unaccent and a pg_trgm similarity tiebreak). A different brand, chassis or part number on both sides vetoes a text match, and so do opposite positions such as front and rear. The agent can also read the database through Tiger's MCP server for questions the CLI does not cover, with only the read-only db_query tool enabled.

What the model does: turn "tengo un spur 68T de TT-02 nuevo, lo doy en 10 dólares" into {"kind":"HAVE","chassis":"TT-02","description":"spur 68T","condition":"new","price_usd":10}, choose the skill, and relay the reply. What it never does: decide who is writing, compute a price, or message someone on its own.

Lessons learned

Four things the real tests changed:

1. The model invented phone numbers. My first skill asked the model to pass the sender's phone. In the first real DM it searched the members table by WhatsApp display name for ten tool calls, then ran the command with --phone 1234567890 and later 593000000000, and told me I was not registered. Hermes already puts the sender of each turn in the environment of the shell it runs, so the CLI now reads it there and ignores any number the model passes. The skills don't mention phones anymore. That also means a member can't ask the bot to act as someone else.
2. The model drifted from the JSON field names. It wrote type instead of kind, price instead of price_usd, explicit nulls and a markdown code fence around the JSON. The values were right, so the validator accepts those variants in code instead of asking the member again.
3. Members should only see answers. Hermes shows tool progress in chats by default ("running rc-listings…"). In a club group that is noise, so the profile turns off tool progress, reasoning and interim messages for WhatsApp; the calls still go to the log.
4. The shop's search is literal. "kit Xray XB2D 2026" found nothing, while "Xray XB2D 2026" found the kit in stock. The quote command now tries up to four rewrites of the member's words (filler words removed, model codes only, brand first) and picks the hit whose title matches best.

A real quote from the build: TAM54500, TT-02 68T spur set, $13.00 at the store, 113 listed as its weight, 0.25 lb × $7.00/lb = $1.74 fee, $14.74 if the member orders it, $16.30 if we order it (+12%, $1.56). Those numbers come from the store JSON and rcparts; the model only passed them on.

Why Does Open Innovation Matter?

Because the business has to cost almost nothing to run. The fee on a $13 part is a couple of dollars. Gemma 4 26B A4B costs $0.0675 per million input tokens on OpenRouter. A model call in this bot is about 20,000 tokens, most of it Hermes's own instructions and tool definitions, so a call is about a seventh of a cent and a full "add a listing" turn is well under one cent. A closed frontier model at many times that price would eat the margin of a club-sized business.

Because I can leave. The weights are open. OpenRouter also lists a free :free version of the same model; during the build its shared pool was rate-limited, so I pay for the regular one. If the price changes or the provider disappears, the same model family runs on Ollama on my own machine and the profile changes one block. I describe that swap; I have not tested it in this build.

Because the agent itself is open. Hermes gave me the WhatsApp bridge, skills, cron and a profile format I can version in git and install on my home server. When something went wrong I could read the session database and the exact tool calls the model made, which is how I found the invented phone numbers.

Because the club's data stays with the club. Listings live in a Postgres database I control. I also looked at self-hosted Cognee, an open-source memory layer, for remembering members' cars and habits. It can run on the same Postgres with open embeddings, but its extraction step makes several model calls per message and it was too much for this week, so it is the next step, not part of this build.

What is not open: WhatsApp itself, and the bridge is unofficial. A version for many clubs needs the official WhatsApp Business API.

My First User

The first user besides me is a fellow racer and member of ACRO (the local RC racing club), who used the bot from their own phone in the test group and in a private chat. They asked for an import quote for a rear anti-roll bar, chose "we order it", sent the deposit receipt as a PDF in the chat, and the order went into the batch after I confirmed it. The receipt was a generated test PDF from a fictional bank, so no money moved in this test.

He was amazed at how the club chatroom history can be leveraged to connect people's spare parts HAVE/NEEDS through an agent that automates this process. The imports service also caught his interest, as club members often opt to make individual imports to cover their needs and end up paying extra fees. If a club starts batching orders and sharing fees, the hobby becomes more accessible.

Prize Categories

Best Use of Gemma. Gemma 4 26B A4B is the only model in the project. It reads every message in the club group and in DMs, turns HAVE and NEED messages in Spanish or English into listing JSON, picks the skill and relays the replies.

Best Use of Tiger Data. The listings, matches, quotes and orders live in a Tiger Cloud Postgres service (Postgres 18, TimescaleDB 2.30). Matching uses Postgres full-text search with unaccent and pg_trgm, driven by a bilingual RC glossary, and the agent has Tiger's MCP server with the read-only db_query tool for questions the CLI does not cover; all writes go through the CLI. pgvector and pg_textsearch (BM25) are available on the same service and are the next step for words the glossary does not cover.

Limits

• The WhatsApp bridge is unofficial, and WhatsApp can restrict the bot's number.
• Inference is at OpenRouter, not local. The self-hosted path is described, not tested.
• The demo club and its listings are seeded test data, and the deposit receipt in the demo is a generated test PDF.
• The store does not say what unit its weight is in. 113 for the spur set is consistent with grams, and the quotes say "weight as listed by the store"; I have not checked it against a scale yet.
• One online hobby shop I tried to include in the prototype disallows search in robots.txt.
• No payments: quotes and orders are records, and I collect and place orders by hand. ACP, UCP, x402 protocols can be explored in the future.
• A turn takes 20 to 60 seconds, mostly model calls on a 20,000-token prompt.
• Matching is words, not meaning. Spanish and English meet through a 42-concept RC glossary; words outside it will not match across languages until embeddings are added.