Paper to Portfolio With Local Gemma

No, local AI isn't required. Week 3 asks for open-source AI at the core, and any one of these qualifies:

• an open-weight model
• an open-source agent harness or framework
• local inference

Or all three. The rule is that the open pieces have to be what makes the project work. Running on your own machine is just one way to meet it.

That's why so many of the categories are hosted services:

• Hosting an open model somewhere else still counts. Gemma "served through Google Cloud" and an open model on a DigitalOcean GPU Droplet are both still open-weight models.
• Most partners are supporting pieces, not the AI core. Render (hosting), MongoDB or Tiger Data (storage), Sentry (tracing), Temporal (durability), Auth0, SerpApi and .tech domains all sit around an open model. They don't replace it.
• Partner categories are optional. "Use of Partner Technology" is the last judging criterion and it's marked optional. The main prize doesn't need any partner.
• I can't confirm whether some are open. Ultramarine is reached through the OpenAI and Anthropic SDKs. I don't know if its weights are open, so using it as the main model could go against the prompt. Read their docs before relying on it.

For the folder tidier, I suggest switching from qwen3.5:4b to gemma4:e4b, which you already have. That makes the project eligible for Best Use of Gemma ($200, a featured category), and it stays fully offline, which suits a tool that reads your private files. Gemma is in the same Ollama setup, so the build doesn't change. It's a bit slower (6.6 GB against 3.4 GB), which is fine for a tool you run once and leave.

Two optional categories are cheap to add later:

• Entire: share the agent session in the write-up.
• GitHub Copilot: a GitHub Actions workflow that runs the tests.

Shall I go ahead with the folder tidier in Python on gemma4:e4b?

Read-it-later killer is shortlisted. More ideas, all built to stack categories:

1. Read-it-later killer ✅ shortlisted Overnight agent reads your link backlog and gives you a 5-minute morning paper plus audio. Categories: Gemma, Tiger, ElevenLabs, Temporal, SerpApi, Copilot, Entire, Render, .tech

2. Patient Buyer You tell it "I want this phone or flight, no rush." It checks prices slowly for days or weeks, and TabPFN predicts whether the price will drop next week, so it says wait or buy now. The theme fits literally: patience pays. The workflow runs for weeks, which is exactly what Temporal is for. Categories: TabPFN, Temporal, SerpApi, Browserbase, Gemma, Copilot, Entire, Render, Sentry, .tech

3. Forward Whisperer (India angle) Family WhatsApp forward comes in. The agent checks the claims with web search and drafts a polite, respectful reply for your uncle in Hindi or English, optionally as a voice note. Repeat forwards are matched against claims it already debunked. It saves you the effort of arguing with relatives. Fun story, very original. Categories: Gemma (multilingual), SerpApi, ElevenLabs, Tiger (vector match), Temporal, Copilot, Entire, Render, .tech, Sentry

4. Repo Gardener Slow, steady help for maintainers. Each night it triages issues, labels them, nudges stale PRs and writes a weekly changelog. Pitch: one careful action per night instead of a flood of AI spam. That ties straight into why Hacktoberfest 2026 stopped counting PRs, so judges will connect with it. Categories: Mastra (lists "triages GitHub issues" by name), Copilot, Gemma, Temporal, Sentry, Entire, Render

5. Screenshot Archaeologist Your 4,000 phone screenshots get captioned and OCR'd overnight by Gemma vision, then become searchable ("wifi password screenshot", "that train ticket"). It also finds duplicates to delete. Categories: Gemma, Tiger (pgvector hybrid search), Temporal, Copilot, Entire, Render, .tech

6. Ramble Journal You record voice notes all day. Overnight they're transcribed locally and Gemma writes a calm evening journal, plus a weekly "letter to future you" read back in a calm voice. You can ask it things like "what was I stressed about last month?" This one covers the relax side of the theme. Categories: Gemma, ElevenLabs, Tiger (pgvector memory), Temporal, Copilot, Entire, Render

7. Reply Tomorrow An inbox agent that never replies instantly. Overnight it drafts replies in your voice, using a model fine-tuned on your sent mail with Tinker and measured against a baseline. Auth0 asks you to approve before anything is sent. Categories: Gemma, Tinker, Auth0, Temporal, Copilot, Entire, Sentry

My picks: 3 is the most original and the best story for the post, which matters because writing quality is weighted highest. 2 is the strongest technical fit, with TabPFN and Temporal used for exactly what they're built for.

Reply with your top 3.

Top 3 saved: Forward Whisperer, Read-it-later killer, Screenshot Archaeologist.

Forward Whisperer

The problem: an uncle posts "NASA confirms Diwali is visible from space, share before midnight" in the family group. You either let it slide, or spend 20 minutes searching, writing a careful reply and keeping the peace. Every week.

What it does: you forward the message to the tool, either as text or a screenshot. It:

1. Reads it. Gemma vision OCRs the screenshot and handles Hindi, English and Hinglish.
2. Splits it into checkable claims, e.g. "NASA said X" or "drinking hot water cures Y". Opinions and greetings are skipped.
3. Checks memory first. Tiger pgvector looks for the same or a reworded claim it has already debunked. These forwards repeat a lot, so cached matches return instantly.
4. Searches the web with SerpApi for each new claim, preferring fact-check sites (Alt News, Boom, PIB Fact Check, WHO) and official sources.
5. Gives a verdict per claim: False, Misleading, True, or Can't verify. Each one links its source. Gemma has to quote the search results and isn't allowed to rely on its own memory.
6. Drafts the reply in your chosen tone and language: respectful ("Chachaji, interesting! I checked, though..."), short, or funny. You copy and paste it. It never auto-sends.
7. Optional voice note. ElevenLabs reads the reply in Hindi, because elders often listen to voice notes more than they read text.

Fit with the theme:

• Saves effort: about 20 minutes of arguing becomes about 20 seconds. Measured, that's the "Lazy Payoff" section of the post.
• Relax: fewer family fights.
• Slow and steady: it's the patient, calm reply you'd write if you had the time and energy. A cooled-down second thought, not a hot-take reply. Optional overnight mode works through the whole family group export and leaves a morning digest of "3 forwards worth replying to, drafts ready".

Category hooks:

Demo: paste a screenshot of a classic fake forward, watch the claims get checked live, get the reply plus voice note. Then paste a reworded version of the same forward and it matches memory instantly. That's very watchable.

Risks and how to handle them:

• The small model gets a fact wrong. It may only judge from the search results, each verdict shows its source, and it defaults to "Can't verify". The post can say so openly, which makes for good writing.
• Politics and religion forwards are sensitive. The tone stays respectful, it never mocks the sender, and it only judges factual claims, not beliefs.
• WhatsApp has no personal API. Input is a paste, a screenshot upload, or a .txt chat export. That's simple and honest.
• Accounts: SerpApi, ElevenLabs, Tiger, Sentry and Render each need a free-tier signup from you. The ElevenLabs connector here also needs authorizing in your claude.ai settings. Core works offline (Gemma plus local search fallback), and each extra is wired in once you have its key.

Scope for tonight: the core loop is doable in about 4–5 hours (OCR, claims, search, verdict, reply, web UI). Then add one integration per hour.

Want me to explain 1 and 5 the same way before you pick, or go with this one?

Yes. All three of your picks have already been built, and so has Repo Gardener. Patient Buyer is the only one with no Week 3 entry so far.

Forward Whisperer: taken twice in this Hacktoberfest, plus many repos

• Forward Check - I built Dhroov an offline bouncer for the family WhatsApp group by forenche, an HF26 Weekend entry: offline Gemma, verdict plus a polite reply.
• Is This Real? A Forward Checker I Built for My Parents by anishisbusy, also an HF26 Weekend entry.
• Aakash-M-z/fwdcheck on GitHub is almost exactly my design: SerpApi, Indian fact-checkers ranked first, a polite reply in the forward's language. Similar repos: tejalp68/Forward-fact-checker-, Ojash-creator/ForwardCheck, sach-ya-jhooth.
• Perplexity and a Meta-backed alliance also run WhatsApp fact-check bots (Meta MCA helpline, Perplexity bot guide).

Read-it-later killer: taken in Week 3, on the exact same stack

• I let an open model delete 1,242 of my saved articles. On purpose. by sarvar_04: TabPFN plus Gemma, local.
• Sloth Brief: an AI news agent that reads overnight by kami_butt_69b922ade8de012, and Reel Sloth by kuntalgorai.

Screenshot Archaeologist: taken in Week 3 and in Week 2

• I let a local Gemma 4 name my screenshots by devcarlat800, a Week 3 entry.
• Memora - Find Your Lost Screenshots by ayesha_rahman_3f84433d77b in Week 2, and Screen Memory by mirshah12 in the Weekend round.

Repo Gardener: taken several times in Week 3

• Leave It: I fine-tuned a 4B model to defer 73 of 731 GitHub issues by memeshe: Tinker plus Temporal, a strong entry.
• I Built an AI Gardener for the Boring Work Every Open-Source Maintainer Avoids by ali_hamza_4bd92ee9f67e78702, and Snooze Agent by ayushg_15.

Week 3 overall is packed with "sloth" overnight digests, file tidiers and group-chat summarizers.

Patient Buyer looks open. The closest earlier project is Anti-Salesman by himanshuc3, from an older, unrelated challenge. It also has the most literal theme fit ("patience pays"), and it uses TabPFN and Temporal for what each is built for. Both are featured or partner categories, and two strong Week 3 entries (Leave It, SlothSRE) show judges are getting partner-heavy builds.

Should I explain Patient Buyer in detail, or come up with a fresh batch of ideas and check each one before showing you?

Checked 13 more ideas. Week 3 is packed: Khata ledgers, family tech support, recipe voice notes, letter readers, pantry/grocery tools, electricity bills, form fillers and money trackers have all been built in Week 2 or 3, or in the Weekend round.

Four ideas came back with no Hacktoberfest entry:

1. Sleep On It (Patient Buyer merged with an impulse cool-off). Recommended. Paste a product link or your cart. Instead of buying, it goes into a cooling-off period that lasts 2–7 days. Meanwhile the agent:

• tracks the price daily
• uses TabPFN to predict the chance of a price drop (e.g. before Big Billion Days or Diwali sales)
• uses TabPFN again to predict your regret chance, trained on your past buys marked "used it" or "regretted it"

When the timer ends it asks: buy now / wait for sale / skip. A running counter shows "₹ saved by being slow". The theme fits literally, since slowness is the feature. The timer that lasts days and survives crashes is exactly what Temporal is for. There was no match for cool-off agents anywhere on DEV. Categories: TabPFN, Temporal, SerpApi, Gemma, Browserbase, Sentry, Render, .tech, Copilot, Entire, plus Tiger or MongoDB for price history

2. Waitlist Oracle "WL 37, will it confirm?" TabPFN predicts the chance a train waitlist ticket gets confirmed. A patient watcher checks the PNR, and if the chance drops it suggests alternatives (Tatkal, other trains, bus). A very Indian pain point. Risk: there's no free source of historical waitlist data, so I'd need a public dataset or synthetic data, which judges may notice.

3. Rent Radar Overnight flat-hunting agent. It scans listings, TabPFN estimates a fair rent so overpriced flats get flagged, and Gemma spots red flags (10-month deposit, "no bachelors") and drafts the broker message. Risk: scraping 99acres or NoBroker may break their terms. A demo on a sample dataset is safer.

4. Khatam Alert ("about to run out") Predicts when the LPG cylinder, water cans and similar will run out, from your booking SMS history, and reminds you before cooking stops. Risk: probably too small to stand out as "cool".

Sleep On It is the strongest: open, buildable tonight, best theme story, and it hits both featured partners (TabPFN and Temporal). Pick one, or pick a top 3 if you want to narrow it down the way you did before.

Found two stock ideas with no existing AI tool. Neither turned up in Hacktoberfest, DEV or GitHub.

1. Dadaji's Shares (my pick)

Old physical share certificates turned into today's value and a claim plan.

Lots of Indian families have a drawer of 1980s share certificates from a grandparent. Nobody knows what they're worth, because the company has since been renamed, merged, split, or given bonus shares five times, and the shares may have been moved to IEPF, the government fund that holds unclaimed shares. Today families either pay recovery firms or spend weeks tracing it. My search found only paid services (Shares Recover, KMFSL, Rook Capital's guide) and no AI tool.

How it works:

1. Snap the certificate. Gemma vision reads the company, folio number, share count, date and holder name.
2. Trace the lineage. An agent searches the web (SerpApi) for renames, mergers, demergers, bonus issues and splits, and builds a timeline like "1987: 100 shares of X → 1992 1:1 bonus → 2004 merged into Y at 3:2 → 2010 2-for-1 split → today 1,200 shares of Y". Every step cites its source.
3. Today's value at the current price, plus dividends likely sitting unclaimed.
4. Claim plan. Checks whether the shares are likely with IEPF, then builds a personal checklist: IEPF-5 form, transmission documents if the holder has died, demat conversion, the right registrar.
5. Explain it to the family. A short voice summary in Hindi.

Theme fit is perfect: shares held 35 years is the ultimate slow and steady, and the tortoise wins. Writing quality counts most, and "Dadaji bought 100 shares in 1987; here's what they're worth now" is a story people will click. Demo safety: validate against famous public histories (Wipro and Infosys bonus/split chains) so judges can check the math. Categories: Gemma (vision plus reasoning), SerpApi, Temporal (long research run with retries), Tiger (corporate-action timeline plus pgvector), ElevenLabs, Sentry, Render, .tech, Copilot, Entire

2. Shareholder Sahayak (helper)

A personal proxy advisor for retail investors. Every company you hold sends AGM and postal-ballot notices: 80 pages about director appointments, related-party deals, executive pay, auditor changes. Retail investors ignore them. The agent reads each notice for your holdings, explains every resolution in plain Hindi or English, flags governance red flags, and suggests a vote with reasons. Institutions pay proxy advisors (IiAS, ISS) for this. I found nothing open-source for Indian e-voting (Harvard corp-gov piece; the only retail product is Proxinsight, in Africa). Weaker on theme and harder to make relatable.

Already taken

Stock research assistants (StockPilot AI by sandeela, Week 3), Trading Bot That Refuses to Trade by Dabhi Rohit (Week 3), trading journals and revenge-trade detectors (many), promise trackers (Say vs Do, Marvin Labs).

Dadaji's Shares or Shareholder Sahayak?

Locked in. Name options with no family reference:

• Kachhua (कछुआ, "tortoise"). Ties straight to the theme: the shares that sat still for 35 years won the race. My pick.
• Dusty Folio. The forgotten certificate in a drawer.
• FolioTrail. Follows a share through every rename and merger.
• Paper to Portfolio. Says plainly what it does.

Which categories fit

Strong fit, real use (8):

Fit with some extra work (4):

Skip (8):

• TabPFN: no real prediction task here, so it would look forced.
• Tinker: no time or data to fine-tune.
• Arduino: needs hardware.
• Mastra: a TypeScript framework, and we're using Python.
• MongoDB: duplicates Tiger. Pick one.
• Backboard: at best a model comparison, so it would look tacked on.
• Ultramarine: a hosted model that could undercut the "open AI at the core" rule.
• DigitalOcean: possible instead of Render, but GPU Droplets cost money.

My suggestion is the 8 strong ones plus Render and ElevenLabs, 10 in total. Every one of those is a real use. Forced integrations can hurt the Technical Execution and Writing scores, which matter more than any single partner prize.

Accounts you'd need to sign up for: SerpApi, Temporal (runs locally, no account), Tiger Cloud, Sentry, Render, ElevenLabs, the .tech claim, and Entire. I'll check whether Entire needs its own tool. Without keys, the core still runs on Gemma plus local search, and each integration switches on once its key is added.

Which name, and are you happy with those 10 categories?

Here's where you stand on the 10 planned categories. I checked your old project .env files (variable names only, not the key values), your offers notes, and your Tiger and MLH accounts.

Ready (4)

• Gemma: nothing to sign up for. gemma4:e4b is already on your machine in Ollama.
• ElevenLabs: key already in touch-grass/.env, side-hustle/server/.env and build-for-a-friend/.env.
• Tiger Data: account connected. You have one service, "touchgrass" (paused). We can reuse it or create a new one.
• Temporal: runs locally with no account. I just need to install the CLI.

Probably ready, please confirm (3)

• Render: code claimed Oct 3, but your notes say RENDER_API_KEY was still empty. Did you redeem it?
• .tech domain: your notes say not claimed yet (code expires Dec 30). MLH shows 7 claims in total, but your notes only list 4. The 3 extra claims are from Oct 8, 13 and 20, and the tool doesn't show sponsor names. Check hacktoberfest.com/my/promos to see whether .tech is one of them.
• GitHub Copilot: your GitHub account is enough (Copilot has a free tier), but the gh CLI isn't installed on this machine.

Not signed up yet (3)

• SerpApi: no key anywhere. The free plan gives 100 searches a month, enough for the demo. Sign up at serpapi.com.
• Sentry: no key. The free tier is fine. Sign up at sentry.io and create a Python project to get a DSN.
• Entire: no sign of it. I'll check what their category needs (probably their session-sharing tool) and let you know.

Unrelated to this build: your $25 DigitalOcean credit unlocks at the Ranchi Hack Day on Oct 28.

Your to-do (about 10 minutes): sign up for SerpApi and Sentry, and confirm Render and .tech. Meanwhile I'll start building the core (certificate reading, lineage tracing, valuation), which needs no keys. Shall I start?

DigitalOcean: not unlocked, and it probably can't be now.

• The credit only unlocks after organizers check you in at a Fest.
• MLH has the Hacktoberfest Hack Day Ranchi on Oct 24, not Oct 28 as your notes say. It's now over.
• Your status there is still "registered", not "checked in".
• The Ranchi event offers nothing now, and the DO pool doesn't show on the main Hacktoberfest event either.

To check yourself: hacktoberfest.com/my/promos shows the DigitalOcean card as locked or "Get your code". If you went to Ranchi but weren't checked in, email hacktoberfest@mlh.io and ask them to mark your check-in. DO links expire Oct 31.

Deadline check (important): MLH's own listing shows Week 3 as already ended at Oct 25, 06:59 UTC. The official schedule feed says Week 3 runs until Oct 26, 06:59 UTC = Mon 12:29 PM IST, and the DEV page says "Submissions due: October 25" (US time). Two sources agree, so we have about 11 hours from now (12:50 AM IST). The MLH listing looks wrong, but don't leave the submission to the last hour.

Given the time, I suggest starting the build now, with a smaller category list: Gemma, SerpApi, Temporal, Tiger, Sentry, ElevenLabs, Copilot, Entire, plus Render and .tech if they're quick. Should I start?

Entire is set up and this planning session is imported: 13 turns, stored locally for now.

You need to do two things:

1. Log in. It opens a browser for GitHub sign-in, so it has to be you. Type this in the Claude Code prompt:

   Then run ! entire import claude-code again so the history syncs to your Entire dashboard. Right now it's local only.

2. Before the first push, a privacy note. Checkpoints upload on git push, and your repo Rahullkumr/slow-and-steady is public, so anyone can read the transcripts. This chat contains your MLH user ID, the Tiger service ID, local file paths and .env variable names. No key values, and I never printed any. If that's fine, push as normal. If not, I can send checkpoints to a separate private repo (entire configure --local --checkpoint-remote github:Rahullkumr/<private-repo>), but then judges can't open the link.

From here on, Entire's hooks record this and every future session automatically. The DEV post's "My Agent Session" section links to https://entire.io/gh/Rahullkumr/slow-and-steady/commit/<sha>.

Two new folders are untracked: .entire/ and .claude/, which holds Entire's 10 hooks. They go in with the first commit, which I'll make when you say so.

Next session: open this repo and say "continue Paper to Portfolio". Memory has the idea, design, categories, deadline (Mon 12:29 PM IST), account status and rejected ideas.

Sources: Entire quickstart, hack.entire.io, import PR #1527, HF26 Challenge Hub.

The server routes work. Waiting on the full Gemma run (extraction plus summary, a few minutes on CPU) to confirm the end-to-end result.

Build done, and the DEV post draft is ready at: C:\Users\rahul\AppData\Local\Temp\claude\c--Users-rahul-Desktop-Hacktober-Fest-slow-and-steady\1655668e-39a6-42b4-b787-938ca8b2660d\scratchpad\dev-post.md

What's built:

• p2p/ holds the Gemma client, the checks, the arithmetic, the claim checklist, the server and a one-page UI.
• 8 unit tests, all passing.
• GitHub Actions CI, a README, an MIT license, and fictional sample data.
• Committed locally as 83893dd. Nothing pushed yet.

About 3 hours 40 minutes left. What I need from you:

1. Entire login, then push. Run ! entire login in the prompt. Then tell me "push" and I'll push the code plus the 14 session checkpoints. Reminder: the repo is public, so the transcripts become public too (no keys in them).
2. Record the demo (about 2 minutes, speed up the waits). Run python -m p2p.server, open http://localhost:8765, click Use fictional sample, then Read certificate, then Trace to today. A full run takes about 5 minutes on CPU, so trim the waiting in the edit, or say so in the video, since slow is the theme. Upload it to YouTube or Loom.
3. The DEV post. Should I create it on DEV as an unpublished draft through DevRelay? I'd fill in the Entire link after the push, and you'd only add the video link and press publish. Or paste it yourself from the file.

Reply with your answers, e.g. "push + draft".

Open this link and approve:

https://in.auth.entire.io/cli/auth?user_code=NFJT-DWVH

Device code: NFJT-DWVH. Sign in with GitHub when it asks.

The login finishes automatically once you approve. Tell me "done", plus whether to push and create the DEV draft.

Logged in to Entire as github:Rahullkumr, and the session is synced (17 turns).

Two decisions are still open:

1. Push? This uploads the code plus the agent transcripts to the public repo Rahullkumr/slow-and-steady. No keys are in them, but your MLH user ID, the Tiger service ID and local file paths are visible.
2. DEV draft? I'd create the post on DEV as an unpublished draft with the Entire link filled in. You add the demo video and press publish.

Reply "push + draft", or say which one. About 3.5 hours left.

Not cleanly. Two facts from Entire's own docs (entire checkpoint delete --help):

1. Every checkpoint carries the session's full transcript. The commit-linked checkpoint 01M5V32K… holds this whole chat from the first message. Deleting the early checkpoints doesn't remove those turns, because they're still inside it.
2. Deleting only removes a checkpoint from the branch tip. The checkpoint branch's history still contains it, so a push would upload it anyway.

So with Entire it's the whole session or nothing. Your realistic options:

• A. Push it all. My pick. The "useless" part is actually the strongest bit for the Entire category, which asks you to "search past agent work to explain why code exists". The transcript shows why this idea exists: we checked 25+ ideas, found them all taken, and verified this one was unique before building. The post can point judges to that story. The cost is that your MLH user ID, the Tiger service ID and local paths are visible. No keys.
• B. Curated session instead of Entire. Turn off session uploads (entire configure --local --skip-push-sessions), push only the code, and save a trimmed transcript to DEV with DevRelay. That one supports picking which turns to keep, and it embeds in the post's "My Agent Session" section. Cleaner, but you lose the Entire category.
• C. Both. Push to Entire anyway for the category, and also embed a curated DevRelay version in the post so readers see the short story. Raw history stays on Entire for anyone who digs.

Which one: A, B or C?

Pushed. The code and all 18 session checkpoints are on GitHub.

• Repo: https://github.com/Rahullkumr/slow-and-steady
• Entire session: https://entire.com/gh/Rahullkumr/slow-and-steady/commit/83893ddd41a6b0a02eb171d1582772a11948174d
• CI: the push triggers GitHub Actions. Check the Actions tab shows the tests passing (green).

The post draft now has the real Entire link. Its "My Agent Session" section turns the long planning chat into a selling point: more than 25 ideas checked, almost all already built, this one the first with no existing tool.

Left to do (about 3h20m):

1. Demo video. Run python -m p2p.server, open localhost:8765, click Use fictional sample, then Read, then Trace. Speed up the waits and upload to YouTube or Loom.
2. DEV post. Should I create it as an unpublished draft on your DEV account? Then you'd only paste the video link where it says VIDEO_URL and press publish. Or copy it yourself from: C:\Users\rahul\AppData\Local\Temp\claude\c--Users-rahul-Desktop-Hacktober-Fest-slow-and-steady\1655668e-39a6-42b4-b787-938ca8b2660d\scratchpad\dev-post.md

Click "Mirror existing GitHub Repository" on the right. Its /gh/... address matches the link in the post.

1. Click it, then authorize or install the Entire GitHub App when GitHub asks.
2. Give it access to Rahullkumr/paper-to-portfolio. "Only select repositories" is fine.
3. Finish, and wait about a minute for it to sync. It reads the checkpoint refs we already pushed.
4. Open https://entire.com/gh/Rahullkumr/paper-to-portfolio/commit/83893ddd41a6b0a02eb171d1582772a11948174d, then the same link in incognito.

Don't use "Create new repository". That hosts the code on Entire under /et/... and the link wouldn't match.

Tell me what incognito shows.

The Entire link is public: incognito shows the commit, the Claude Code (Opus 5.5) session and the full transcript. The Entire category is covered.

Two things left (about 3h10m):

1. Demo video. Run python -m p2p.server, open http://localhost:8765, click Use fictional sample, then Read certificate, then Trace to today. A full run takes about 5 minutes on CPU, so speed up the waits. Upload to YouTube or Loom.
2. DEV post. The draft is ready with the real repo and Entire links. Should I create it on DEV as an unpublished draft now? Then you'd paste the video link where it says VIDEO_URL and press publish. Reply "draft", or copy it yourself from: C:\Users\rahul\AppData\Local\Temp\claude\c--Users-rahul-Desktop-Hacktober-Fest-slow-and-steady\1655668e-39a6-42b4-b787-938ca8b2660d\scratchpad\dev-post.md

The assembler is ready. Waiting for the capture run to finish; Gemma reads, then traces, about 5 minutes in total.

The DEV draft is saved but not published. The demo video is ready too.

DEV draft: https://dev.to/rahullkumr/paper-to-portfolio-a-local-gemma-reads-the-old-share-certificate-in-your-drawer-but-never-does-1gn7-temp-slug-1295761/edit

It has the right tags (devchallenge, hf26challenge, gemma, opensource), AI-assistance disclosure, and the GitHub embed and Entire link filled in.

Demo video: C:\Users\rahul\Desktop\Hacktober Fest\paper-to-portfolio-demo.mp4, 67 seconds, 1080p, captions burned in. The screenshots come from a real run on your laptop: Gemma read the certificate in 38 seconds and traced it in 193 seconds. It opens with a title card and a problem card, then the 3 steps with captions, then an end card with the repo link. A thumbnail is next to it: paper-to-portfolio-thumbnail.png.

To finish (about 3 hours left):

1. Upload the mp4 to YouTube (Unlisted is fine).
2. In the draft, replace the line "▶ DEMO VIDEO: PASTE YOUTUBE LINK HERE" with {% embed https://youtu.be/YOUR_ID %}.
3. Preview, then Publish. Aim well before 12:29 PM IST.
4. Check the repo's Actions tab shows the tests passing (green).

How to run the project yourself:

Then open http://localhost:8765, click Use fictional sample, then Read certificate, then Trace to today. Ollama must be running with gemma4:e4b. A server from the video capture is still running on port 8765, so you can open the link straight away.

The demo video now has ElevenLabs narration, and ElevenLabs is added to the DEV draft's prize categories. The post is still an unpublished draft.

Video: C:\Users\rahul\Desktop\Hacktober Fest\paper-to-portfolio-demo.mp4, 2 minutes 22 seconds, 1080p.

• One voice clip per slide, using your existing ElevenLabs voice (eleven_multilingual_v2). Each slide stays on screen as long as its narration.
• Captions are still burned in, so it works on mute too.
• I checked the file has an audio track at normal volume, but I haven't listened to it. Play it once before uploading.

.env: I copied ELEVENLABS_API_KEY and ELEVENLABS_VOICE_ID from touch-grass/.env into the repo without printing them. .gitignore already lists .env, and git check-ignore confirms it won't be committed.

DEV draft updated:

• A line in the Demo section saying the voice-over is generated with ElevenLabs.
• Best Use of ElevenLabs added, making 4 categories: Gemma, Entire, GitHub Copilot, ElevenLabs.
• Edit link: https://dev.to/rahullkumr/paper-to-portfolio-a-local-gemma-reads-the-old-share-certificate-in-your-drawer-but-never-does-1gn7-temp-slug-1295761/edit

Left for you (about 2 hours to the 12:29 PM IST deadline):

1. Upload the mp4 to YouTube. Unlisted is fine, and paper-to-portfolio-thumbnail.png next to it can be the thumbnail.
2. In the draft, replace "▶ DEMO VIDEO: PASTE YOUTUBE LINK HERE" with {% embed https://youtu.be/YOUR_ID %}.
3. Preview, then Publish.

The new thumbnail is ready and the shadow is clean now.

• For YouTube: C:\Users\rahul\Desktop\Hacktober Fest\paper-to-portfolio-thumbnail-1280.jpg (1280×720, YouTube's size)
• Full size: C:\Users\rahul\Desktop\Hacktober Fest\paper-to-portfolio-thumbnail.png (1920×1080)

The layout:

• The tilted certificate with a shadow on the left; its "SPECIMEN – FICTIONAL SAMPLE" stamp keeps it honest.
• The hook on the right: 1989, 100 shares, then an arrow and "35 yrs", then TODAY, ₹4.6 lakh in big ochre text.
• 11.25x and 1,125 shares below that.
• A Hacktoberfest Week 3 badge at the top, and a bottom strip reading "Paper to Portfolio · Local Gemma · no cloud · no GPU".
• Colours come from the Hacktoberfest brand palette.

You can also upload the PNG as the cover image on the DEV draft, which makes the post card stand out in the challenge feed.

Left (about 2 hours to 12:29 PM IST): upload the video with this thumbnail, paste the YouTube embed into the draft, and publish.

Yes, it's possible, but I wouldn't switch before the deadline.

What works:

• The Gemini API (with an AI Studio key) also serves Gemma models, not just Gemini. Gemma is still open-weight, so the project stays within the challenge prompt.
• The Gemma category explicitly allows serving Gemma "through Google Cloud or another provider".
• It would be much faster: seconds per step instead of 38 seconds plus 3 minutes on your CPU.
• It would allow a hosted live demo (e.g. on Render), since Render's free tier can't run Gemma locally. Judges could click a link instead of installing Ollama.

What you'd lose or risk:

• The main pitch breaks. The post, video and thumbnail all say "the certificate never leaves home · no cloud · no GPU". In API mode, a family's name and folio number go to Google's servers.
• The slow-and-steady angle weakens. The local CPU run is part of the theme story.
• Structured output may not carry over. The code relies on Ollama's strict JSON schema, and the Gemini API may not enforce response schemas for Gemma models the same way. We might need prompt-only JSON plus extra parsing and validation.
• I haven't checked which Gemma versions the API offers right now. We'd need to list the models first.
• Redo work: the code change (30–45 minutes plus testing), then re-recording the video, re-narrating, and updating the post and thumbnail. With about 2 hours left, that's risky.

My recommendation: submit as it is, fully local. Afterwards, add an optional switch (e.g. P2P_BACKEND=gemini, with GEMINI_API_KEY in the gitignored .env) and keep local as the default. Then the post can say "local by default; API mode for a quick demo, and never send a real certificate through it". That gives you a hosted demo without breaking the privacy story.

I can't upload to YouTube from here; there's no YouTube connection. It takes you about 3 minutes, and everything to paste is below.

Steps

1. Go to https://studio.youtube.com, click Create, then Upload videos.
2. Choose C:\Users\rahul\Desktop\Hacktober Fest\paper-to-portfolio-demo.mp4.
3. Paste the title and description below.
4. Thumbnail: upload C:\Users\rahul\Desktop\Hacktober Fest\paper-to-portfolio-thumbnail-1280.jpg. Custom thumbnails need a phone-verified YouTube account.
5. Audience: No, it's not made for kids.
6. Visibility: Unlisted or Public, then Save, then copy the link (https://youtu.be/...).

Title

Description

Tags: hacktoberfest, gemma, ollama, open source ai, local llm, share certificate, IEPF, python

When you have the link, send it to me. I'll put {% embed https://youtu.be/... %} into the DEV draft in place of the placeholder, and then you only need to press Publish. About 2 hours left.

Here's the description with hashtags at the end. YouTube shows the first three above the title.

Still put the plain tags in YouTube's Tags field too: hacktoberfest, gemma, ollama, open source ai, local llm, share certificate, IEPF, python

Send me the YouTube link once it's uploaded and I'll put it into the DEV draft.

The video is in the DEV draft: {% embed https://youtu.be/qGk3mJ9d5nk %} sits in the Demo section, and DEV saved it without complaint. The post is still an unpublished draft.

Open: https://dev.to/rahullkumr/paper-to-portfolio-a-local-gemma-reads-the-old-share-certificate-in-your-drawer-but-never-does-1gn7-temp-slug-1295761/edit

Last steps (about 1h50m to 12:29 PM IST):

1. Optionally upload paper-to-portfolio-thumbnail.png as the cover image.
2. Click Preview: check the video, GitHub card and table render.
3. Click Publish.

Or say "publish" and I'll publish it from here (but then I can't add the cover image).

I've sent the cover image to the DEV draft. DEV accepted the update but doesn't echo the cover back, so check it shows in the preview.

Cover: a wide crop of the thumbnail (2000×840, DEV's cover ratio) showing the certificate and "1989 · 100 shares → TODAY ₹4.6 lakh · 11.25x".

• It's in the repo at docs/images/cover.png, committed and pushed as 387d117. That push also uploaded 1 new Entire checkpoint.
• DEV loads it from https://raw.githubusercontent.com/Rahullkumr/paper-to-portfolio/main/docs/images/cover.png, which returns the image.

Last step: open the draft, click Preview to check the cover, video, GitHub card and table, then Publish. About 1h45m to 12:29 PM IST. https://dev.to/rahullkumr/paper-to-portfolio-a-local-gemma-reads-the-old-share-certificate-in-your-drawer-but-never-does-1gn7-temp-slug-1295761/edit

If the cover doesn't show, upload C:\Users\rahul\Desktop\Hacktober Fest\slow-and-steady\docs\images\cover.png with DEV's "Add a cover image" button.

The cover is set now. The earlier update didn't carry it; a direct API call fixed it. DEV confirms:

• cover_image is your cover at 1000×420.
• social_image (the link-preview card) is the same image at 1000×500.

Important: if the editor tab is still open from before, reload it (F5) before you click anything. Saving from an old tab could overwrite the cover with blank.

Then Preview to check the cover, video, GitHub card and table, and Publish. https://dev.to/rahullkumr/paper-to-portfolio-a-local-gemma-reads-the-old-share-certificate-in-your-drawer-but-never-does-1gn7-temp-slug-1295761/edit

About 1h40m to 12:29 PM IST.

Screenshots look right. Committing and pushing the README: