Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 9 min read

Don't Open 15,000 Tabs Before Slush: The Research Agent We Built at Junction 2026

This is a submission for the MLH x DEV Writing Challenge. How many tabs did you open the last time you prepared for a conference? Now picture a business angel preparing for Slush. One calendar. No analyst. 2,989 attend

Don't Open 15,000 Tabs Before Slush: The Research Agent We Built at Junction 2026

This is a submission for the MLH x DEV Writing Challenge.

How many tabs did you open the last time you prepared for a conference?

Now picture a business angel preparing for Slush. One calendar. No analyst. 2,989 attendees on Cala's list. Five tabs per startup: website, LinkedIn, a funding database, search, news.

That's about 15,000 tabs.

And after all that, you get to the only part that matters, the decision, already tired.

We spent 48 hours at Junction 2026 making that number zero. We finished third in Cala's track, Make Unicorns Real.

This post covers three things:

  1. What we built, and the one rule that keeps it honest.
  2. The trick that took us from 5 to 27 trusted startups without adding a single data source.
  3. What breaks when you build a real-time product on a young graph API, and how we worked around it.

What I Built

First point: the product.

Grabbit is a proactive research agent for investors who decide alone.

  • You give it a vague thesis: "AI startups at pre-seed and seed."
  • It asks two or three questions, each backed by numbers from the candidate set.
  • You name the investors you already trust.
  • It walks their funding rounds in Cala's graph, keeps the startups attending Slush, and ranks them by how close they sit to the people you respect.
  • You correct the report in plain words, "shorter, one line per startup", and it rewrites itself.
  • When you leave, it keeps watching. If a startup on your shortlist moves, you get an alert.

Grabbit never tells you what to buy. The investor decides.

Here's the whole idea seen from the investor's side:

What the investor wants What usually gets in the way What Grabbit does
A shortlist worth their time Too many candidates (2,989) Ranks by proximity to investors they already trust
To believe the numbers Data with no clear source A source badge on every fact, from Cala's source_rank
To get there fast Hours of manual research A first report from a single chat turn, then alerts
To do less work ~15,000 tabs No tabs. You talk to the report

Second point: where it came from.

It didn't start with a stack. It started with two investor interviews, kept anonymous in our research notes. One of them summed it up:

"I have no time to scout; if a tool pre-selects and I review each pick by hand, I've already won."

Junction's Cala challenge, Make Unicorns Real, asked for a product that solves a real need for founders or investors, powered by reliable public data. Slush started right after the hackathon. The brief said to aim for something they could start using the day after. That deadline became our constraint.

Third point: the team.

I brought the first idea, ran the interviews, sat in the Cala briefings, coordinated the team, built the web app and cut the pitch film. I didn't do it alone:

  • JosΓ© Javier Bravo Sarabia built the discovery engine (the reverse walk, the second ring of co-investors, the proximity score) plus the shared cache, the PWA and the deploy.
  • Juan LΓ³pez Bravo owned the product flow and research docs, the Cala client and cache, report polish, the guided demo and the peer-review script.
  • RubΓ©n Moyano PalazΓ³n shaped the agent's questions and focus, added live voice transcription, and ran the engine review and data audit.

Four people, each driving AI coding agents on separate branches. More than 100 reviewed pull requests. One weekend.

The two lessons I'm keeping

Lesson 1: if the question lives in a graph, don't read the big set in list order.

The naive way: go through Slush attendee by attendee. Cost: about 9,000 Cala calls, and you still miss most of the interesting companies.

The way that worked: start from the investors you trust, walk their rounds, then intersect with the event.

The result on our real cache, with Y Combinator and 500 Global as the trusted pair:

  • Startups with a trusted investor in their rounds: 5 β†’ 27.
  • 8 of the top 10 in our ranking were invisible before.
  • Reproducible from the repo.

Same data. Different direction. Five times the signal.

Lesson 2: an agent that invents a number is worse than no agent.

So the model never touches the numbers:

  • Ranking is deterministic.
  • Gemini writes only prose with placeholders.
  • The server fills every value from a sourced fact.
  • A missing field reads "Not published", never zero.

Trust has to be visible on the fact itself, not in a footnote at the end.

Demo

Try it in four steps, about a minute:

  1. Tap "AI startups at pre-seed and seed". Answer or skip the questions.

Grabbit screenshot 1
Grabbit checks the Slush attendees first, then asks only what the data can't settle: which investors you trust. (Guided demo, sample data.)

  1. Type "I trust Y Combinator and 500 Global". The report opens, ranked by paths to those investors, with a source on every fact and a "Why it is here" path through the graph.

Grabbit screenshot 2
Left: the path Grabbit walked in Cala's entity graph, step by step. Right: the report it wrote from it. (Guided demo, sample data.)

  1. Type "Shorter, one line per startup". Watch it rewrite itself. The history keeps what changed, with undo.
  2. Press Alerts. About thirty seconds later, an alert lands.

Grabbit screenshot 3
The loop closes: a startup on your shortlist moves and you hear about it. Labelled "Simulated event" on purpose. (Guided demo, sample data.)

One honest detail, because I'd want to read it in someone else's write-up: the demo alert is staged, and it's labelled Simulated event on screen. No startup raises money on cue. It still goes through the real diff, the importance check and the delivery path.

The deployed API answers from a cached copy of Cala (CALA_MODE=replay) so the demo survives venue wifi. Y Combinator and 500 Global are the warmed pair.

Want to run it yourself?

  • No keys: replay mode plus scripted model answers.
  • With keys: set CALA_API_KEY and GEMINI_API_KEY and warm the cache once.

Partner Technologies

Take Cala out of Grabbit and nothing is left.

The question Grabbit answers, which startups at Slush are a few funding rounds away from the investors I trust?, lives in Cala's funding-round graph: who invested, who led, when, and according to which source.

How we used Cala

We built on the raw entity endpoints, as Cala asked at the workshop, not on the LLM knowledge search they were retiring. Four calls carry the whole product:

  1. Search entities by name.
  2. Introspect a node.
  3. Retrieve a node and its neighbours.
  4. Knowledge query, only for news.

On 15 November our cache held 1,145 Cala responses: 484 funding rounds, 226 companies with properties, and the 2,989 Slush attendees.

What was great

  • Rounds come with participants, roles and is_lead.
  • Every fact can carry source_rank and an honest date precision.
  • Events are entities, so we could start from the real attendee list.

What hurt, and what we did about it

The data is strong. As a platform for a real-time product, it's still young.

  • No subgraph endpoint. A first chat turn can mean hundreds of calls: 898 in one measurement, mostly served from cache.
  • Undocumented rate limit. Four parallel requests tripped a 429 with no Retry-After. A single serial burst tripped it after about twelve calls.
    • Fix: every request goes through one wrapper. One request in flight, 250 ms spacing, backoff, and a cache keyed by method, path and body, 404s included.
  • Outages. On Saturday around 11:40 we saw 403s with an HTML body for several minutes. Fix: we served from cache until Cala came back.
  • One fund, many entities. A search for Y Combinator returned ten hits: Company, Organization, Group, "Y Combinator W24"… most with no rounds. Fix: a single fund-identity rule by id or name, used everywhere, because matching on identifiers alone wasn't enough.
  • An edge you can't trust. The direction of INVESTED_IN was inconsistent. Fix: we count only round participants.

The rest of the stack

Around that core we used four more partner technologies, and we entered each of their tech challenges. They help, but they aren't the product:

  • Gemini (Best Use of Gemini API) for chat with tool calls, self-rewriting report instructions, report prose, importance scores and live voice transcription. The model talks; it doesn't invent figures.
  • Aiven (Best Use of Aiven): Aiven Runtime hosts the API, and Aiven for Valkey is the shared Cala cache, so warming the graph once served both our laptops and the deploy.
  • Context.dev (Best Use of Context.dev) for unverified web signals, always labelled "web, unverified" and kept separate from Cala facts.
  • ElevenLabs (Best Use of ElevenLabs) for the sound of our pitch film: the orchestral music track, the sound effects on every cut, and the voiceover over the app captures. Not in the product. In the story we told about it.

Hackathon Experience

Junction 2026: European Tech Renaissance. 13–15 November, Hype Areena, Espoo, Finland. About 2,000 builders, 48 hours, and partner challenges instead of a single prompt.

We went as a team of four (JosΓ© Javier, Juan, RubΓ©n and me) for the Cala track, Make Unicorns Real.

We're not a team that met on Discord the week before. JosΓ© Javier, Juan and RubΓ©n are former coworkers and people I knew from university. Above all, they're close friends. That's why we went together, and that's why we had a great time doing it.

The weekend, in four beats:

  • Friday: we committed to Cala as the verified source and to the raw entity API.
  • Saturday morning: the persona locked in. A business angel, Slush as the starting list, a trust badge on every fact.
  • Saturday night: the reverse walk landed, and suddenly the shortlist made sense.
  • Sunday: review, fixes, a data audit on the real cache, the pitch video, and credit for the work that never shows up in git log: the idea, the interviews, the briefings, the coordination.

The GitHub history is honest about the hour. My last commits on the pitch film and the team section landed after 02:00 on Sunday, Madrid time. Juan was still fixing the composer sizing in Safari around 03:00.

I won't invent a dramatic 3 a.m. outage I can't point to. The real breakage was slower and more useful:

  • an API that's excellent at truth and awkward at volume,
  • a fund that turns out to be ten entities,
  • venue wifi you shouldn't bet a live demo on,
  • and a team that kept reviewing pull requests instead of merging on hope.

Then Sunday's ceremony: third place in the Cala track.

But the result isn't what I'll remember first. It's the arena. Hype Areena was packed with talent, and the atmosphere was incredible. My favourite part: meeting so many Spaniards who, like us, took the risk of travelling that far to build for a weekend.

Want to go deeper?

So, back to the question I started with: how many tabs did you open the last time you prepared for a conference?

With Grabbit the answer is zero. You talk to the report instead.

Third place in the Cala track says the idea holds up. You decide whether the product does.

If you're an investor who decides alone, or a builder who has ever drowned in tabs the night before a conference: the demo is live. Try it, then tell us in the comments where it's wrong. That's the point of the agent: the more you talk to it, the shorter and more precise it gets.

JosΓ© Javier, Juan, RubΓ©n: see you in the comments.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.