Harshit DuggalWorkAboutDark FunnelBlog
Harshit Duggal
Work with me

Type a company. Get profitability, burn, and headcount from public sources.

I built Bixbite to get a quick read on what public sources can actually support about a company. Enter a name or domain to see profitability, burn, headcount trends, and an IPO or acquisition outlook. Every claim links to its source. If the evidence runs out, the answer stays unknown.

Public-data productEvidence pipelineSide project
bixbite.darkfunnel.pro

Summary

Name a company. Get a sourced verdict.

Bixbite is a lookup. Type a name or domain. It returns profitability, burn, headcount trend, and IPO or acquisition outlook. Every claim links to its source. If nothing public exists, it stays unknown.

I built it as a side project to solve my own problem, then shared it on Reddit and X.

The part I kept coming back to was deciding what deserved to appear as a result. I wanted the lookup to be useful without making thin evidence look complete:

  • Every claim needs a source you can open.
  • Missing evidence stays unknown instead of becoming a guess.
1,120
searches run
150+
anonymous users, no signup
500+
distinct companies looked up
1.6 s
evidence budget per search. Cache hits skip it.

One lookup

Give it Gong.

Gong is a useful example because not every number has the same support. Bixbite calls profitability "Doing okay," headcount "Stable," and burn "Moderate." It also lists ARR at $500M and marks it not verified. That label belongs beside the number, so you can decide how much weight to give it.

I read the result in two parts:

  • The labels summarize what the available sources support.
  • The ARR figure is present, but it is not verified.
SignalResultRead
Profitability
score 51
Doing okayARR $500M. Marked not verified.
Headcount
score 58
Stable2,114 people, down 1.6%. No layoffs on record.
Burn
score 57
ModerateRevenue growing.
Outlook
IPO possible, acquisition unlikelyMarked not verified.
Gong, 26 October 2026.

Where the data comes from

Public sources, read as a waterfall.

LinkedIn company pages, news, filings, funding announcements, and the company’s own site. Nothing private. Nothing behind a login.

Finding a source is only half the job. Bixbite also needs to tie it to the right company and decide whether the figure can support an answer:

  • A company record must match the company’s website.
  • The judge checks each candidate figure.
  • A weak match or unsupported figure stays out of the result.
TierWhat it doesWhen it fails
Search
eight buckets at once Shipped
News, people, identity, financials, filings, exits, workforce, and the typed domain. 1.6 s each.A slow bucket is dropped, not awaited.
Enrich
records + LinkedIn Shipped
Structured company records and the LinkedIn page, from cache. 0.9 s.A record must match the company’s website or it’s ignored.
Recover
gaps only Shipped
Re-runs only what came back empty.Still empty stays unknown.
Judge
GPT-6 Luna Shipped
One typed call classifies the bands and checks each candidate figure.No passing judgment, no figure.
Verdict
plain code Shipped
Combines headcount band, confidence and profitability.The verdict is code, not a model’s opinion.
Each tier and its failure mode.

Architecture

Cache first. Search in parallel. Fill gaps. Otherwise unknown.

A hit comes from cache: no research, no quota. A miss fans out to eight searches at once, enriches from company records and LinkedIn, retries what’s empty, then makes one typed judgment. Whatever is still missing stays null.

Where the speed comes from

Search buckets run together, so their times don't stack. Identity, cache, quota, and abuse checks share one database read. One judgment call checks the candidate figures, and saving the result happens after the response. That keeps the save from holding up the lookup.

  • No sequential hops. Search buckets run together. The slowest one sets the pace, not the sum.
  • One database read. Identity, cache, quota and abuse checks go out as one batch.
  • One judgment call. Company bands and every figure check share a single round trip.
  • Write after respond. Saving the result doesn’t hold up the answer.

That split matters when sources are thin. A cached answer makes repeat lookups cheap, but it doesn't make an incomplete result complete. Anything the evidence can't support stays null.

Judgment

Four calls that shaped it.

  • Public data only. No accounts, no paid databases. Private companies get thin evidence.
  • Unknown beats a guess. Five typed result states. An outage never reads as “no data.”
  • A model judges, code decides. GPT-6 Luna rates the evidence. The verdict is deterministic.
  • No signup. An anonymous cookie, five free searches per six hours, cached replays free.

I use public data only. There are no accounts or paid databases to fill in the gaps, so private companies can have thin evidence. When a figure doesn't pass the evidence check, Bixbite leaves it unknown. Its five result states also keep an outage from looking like a valid "no data" answer. GPT-6 Luna rates the evidence, but plain code applies the bands and builds the verdict.

What I didn’t build

  • A score with no sources.
  • A guess for missing data.
  • Accounts.

There's no signup. An anonymous visitor gets five searches every six hours, and cached replays are free. I wanted someone to be able to check a company without first handing over an email address.