Asli: is that GPU deal real?

Paste a GPU listing or a screenshot. Asli runs NVIDIA Nemotron on Nebius Token Factory to check the seller, real stock and trusted-store prices, and catches bait "Notify Me" deals.

Comment

Inspiration

I stayed up until midnight for a Flipkart RTX 3060 listed at ₹28,839, 66% off. At 12:00 it said the item was not available: "notify me when it comes back". It was not the first time.

So I started watching. On day one, 18 of 22 Flipkart GPU listings I tracked were stuck on Notify Me, and they were the cheap ones. Trusted Indian PC stores list the same RTX 3060 from ₹23,900, but every one of those was out of stock. The cheapest one you can actually buy is about ₹54,662. Buying a GPU in India means wading through bait prices, inflated MRPs and third-party sellers whose warranty may not be honoured.

Asli (असली) is Hindi for "genuine".

What it does

Paste a link from Flipkart, Amazon or an Indian PC store, or upload a screenshot of any listing. Asli streams an agent run you can watch, step by step:

  1. Reads the listing: price, MRP, seller, stock state, returnability. On trusted stores the price and stock are cross-checked against the page's own schema.org data.
  2. Prices the same card at trusted Indian stores (MD Computers, PrimeABGB, Vedant, EliteHubs, The IT Depot, PC Studio) through Tavily, then checks real stock on each product page.
  3. Searches Reddit and TechEnclave for reports about the seller and listing.
  4. Looks up warranty rules for the brand.
  5. Checks watch history: how often this listing was actually buyable.
  6. Weighs the evidence into a verdict: good deal, fair, overpriced, suspicious or likely bait, with a trust score, reasons that link to their sources, and a cheaper in-stock alternative when one exists.

Sites that block servers (Flipkart blocks Render's IPs) are handled by asking for a screenshot. Nemotron 3.5 Super VL reads it in about 4 seconds.

Every checked listing joins a watchlist. A watcher revisits listings every two hours without spending a single model token (it reads page data directly), so a listing that never becomes buyable gets caught with proof.

Real results from the hosted app:

  • Amazon RTX 3060 at ₹58,249: overpriced, trust 45. Third-party seller, non-returnable, above the trusted-store median. The same card is in stock at MD Computers for ₹3,587 less.
  • Flipkart RTX 3060 at ₹28,839: likely bait, trust 15. Notify Me, MRP of ₹86,000, in stock 0 of 8 watch checks.

How we built it

  • Models: NVIDIA Nemotron on Nebius Token Factory, through the OpenAI-compatible API. Each step runs on the model that measured best for it:
    • Read the listing: Nemotron 3 Ultra, reasoning off
    • Read a screenshot: Nemotron 3.5 Super VL
    • Price trusted stores: Nemotron 3 Ultra
    • Summarise forum reports: Nemotron 3 Nano
    • Final verdict: Nemotron 3 Ultra, reasoning off
  • Search: Tavily search and extract, limited to trusted store and community domains, cached and metered against a monthly credit budget.
  • Backend: FastAPI streaming server-sent events, evidence steps in parallel, Postgres on Neon (SQLite locally).
  • Grounding: the model never invents the alternative. The "buy elsewhere" link is picked deterministically from the verified price table, and any source URL not present in the evidence is dropped.
  • Hosting: Render free tier. GitHub Actions keeps it awake and triggers watcher passes.

Choosing models by measurement

I did not guess which model to use. An automatic eval with no hand labels:

  • Reading listings: 29 real trusted-store pages, scored against each store's own schema.org price and stock data.
  • Verdicts: 12 scenarios with rule-defined correct calls.
  • Rule: the cheapest setup within 2 points of the best accuracy wins.
Step Setup Accuracy Cost per call
Read listing Nemotron 3 Nano, reasoning off 84% $0.0002
Read listing Nemotron 3 Nano, reasoning on 88% $0.0005
Read listing Nemotron 3 Super, reasoning off 34% $0.0012
Read listing Nemotron 3 Ultra, reasoning off 98% $0.0034
Verdict Nemotron 3 Nano, reasoning on 75% $0.0005
Verdict Nemotron 3 Super, reasoning on 75% $0.0016
Verdict Nemotron 3 Ultra, reasoning off 100% $0.0022
Verdict Nemotron 3 Ultra, reasoning on 100% $0.0047

Nano got about one price in six wrong, which is too many for a product about prices. Turning Nemotron reasoning off cut output tokens about 7x on extraction with the same answers on Ultra, and halved the cost of the verdict at the same accuracy.

A full check costs about 4 cents.

Challenges we ran into

  • Flipkart blocks datacenter IPs. Rather than fight bot protection, Asli asks the user for a screenshot and reads it with the vision model.
  • Collection pages guess stock. Models (Nano especially) would mark products in stock from a category page. Asli now verifies stock on each product page's structured data.
  • Hallucinated alternatives. An early version recommended a 404 URL on a real store, and once a "cheaper alternative" that cost more. Fixed by grounding the recommendation in the verified price table.
  • Staying alive until January on promo credits. Built a spend ledger from token counts and real prices, a hard total cap, a monthly cap, a Tavily credit budget, per-IP rate limits and a one-hour result cache, so the hosted demo can never run up a bill.

Accomplishments that we're proud of

  • A real problem with real data: every number in the demo comes from live listings.
  • Model routing chosen by an eval, not by vibes, with a visible "Under the hood" page showing accuracy, cost and live usage per step.
  • The watcher catches bait listings for free, with no model tokens.

What we learned

  • Reasoning is not free and not always better. For extraction, Ultra with reasoning off beat every reasoning setup at a fraction of the tokens.
  • Bigger is not always better per dollar, and smaller is not always good enough. Measuring took an afternoon and changed every routing decision.
  • Structured data (schema.org) already on store pages is a great free ground truth for evaluating LLM extraction.

What's next for Asli: is that GPU deal real?

  • Price drop and "actually in stock" alerts for watched listings.
  • More categories: CPUs, SSDs, monitors.
  • A browser extension that checks the listing you are looking at.

Try it

Built With

Share this project:

Updates

Submission history