littlebit.dev
Book a call

Research as a service · Yuriy Babyak

Every company on your list, researched to the source.

I build a research pipeline around the data you need: companies, founders, investors, markets. You get a structured dataset back, with a source behind every field.

Book a callEmail me

Starts with a one-week pilot. If it doesn't beat what you use now, you don't pay.

The question I hear most

Why not just use ChatGPT or Claude deep research?

They answer one question with a report, once. You need the same facts about hundreds of companies, as rows you can check, kept current when something changes.

The benchmark will measure the difference
Output
This pipelineRows in your schema, one per entity
Deep researchA written report
Scale
This pipelineYour whole list, in parallel
Deep researchOne question at a time
Sources
This pipelineA source and confidence level on every field
Deep researchCitations inside the prose
Conflicts
This pipelineResolved, with the reason recorded
Deep researchLeft for you to spot
Updates
This pipelineReruns on a schedule, a news signal or your own signal
Deep researchRun once
Delivery
This pipelineAPI, MCP or a file
Deep researchRead it, copy what you need

How it works

One question goes out to a fleet of agents. A dataset comes back.

Skip this section

Step 1 of 6: Question

Question

Research this VC firm.

Your schema

people[]name, role, focus
sectors[]what the fund invests in
stage[]which rounds it joins
cheque_sizeinferred from investment history
thesisin the fund's own words, with source
portfolio[]company, what it does

The hard part

The web rarely agrees with itself. Most research tools pick one answer and move on.

A fund's website, a news article and a data aggregator all describe the same firm, and they often disagree. A report shows you one version and you never see the others. The pipeline finds every disagreement, decides which value holds, and keeps the rest on record.

2,100+

conflicting values across 100 VC firms

Each dot is one conflict, found and resolved

Question

How much does Northfield Capital invest per deal?

Fund websiteAbout page, 2023

$500k to $2M

Used, but incomplete
NewsFour rounds, 2024 to 2025

$750k, $1.2M, $2M, $2.3M

Used as evidence
Data aggregatorProfile, last updated 2021

Rejected value: $100k to $500k

Rejected, out of date
Kept in your dataset

$500k to $2.3M

Based on real rounds, not the fund's own claim. Every reason is saved.

Northfield Capital is a fictional firm.

Where it pays off

Research once, then keep it running.

Companies change every week. The pipeline reruns the same research on the same rows, so your dataset stays current and every change is marked.

  • On a scheduleEvery few hours, days or weeks, at the pace your data changes.
  • On a news signalA funding round or a new hire starts research on the companies involved.
  • On your signalA new deal in your CRM, an uploaded list or an API call.
One schema, four runsChanged value
The same rows are researched again on each trigger, and changed values stand out.

Benchmark · results pending

Same research tasks. Deep research tools, research APIs and this pipeline.

I'm running the same research tasks through deep research from Claude, ChatGPT and Grok, through research and enrichment tools such as Parallel, Exa and Clay, and through this pipeline. The results go here, with the full method behind them.

Why measure first. An earlier research product I built produced reports clients liked, but $100 per report was too much when ChatGPT covered about half of it.

Results pending. No numbers have been measured yet.

Benchmark of deep research tools, research APIs and this pipeline, results pending
Deep researchResearch and enrichment
Compared onClaudeChatGPTGrokParallelExaClayThis pipeline
Depthfields filled per entityPendingPendingPendingPendingPendingPendingPending
Coverageentities found for the taskPendingPendingPendingPendingPendingPendingPending
Structureoutput usable as dataPendingPendingPendingPendingPendingPendingPending
Sources per fieldvalues that trace to a sourcePendingPendingPendingPendingPendingPendingPending
Cost per entitywhere measured and comparablePendingPendingPendingPendingPendingPendingPending

Published withtask definitions, comparison setup, model versions, evaluation method, sources and limitations. As HTML tables, JSON and CSV.

More tools may be added before publishing. Costs and timings are shown only where measured and comparable.

Email me the results when they're out

At scale

Research at scale, made cheap enough to run on every entity.

5,000
VC firms researched in depth, with 131,000 portfolio companies behind them
$0.73
for a full 437-call investigation of one VC firm
~900
agents running in parallel on own servers

Measured on the VC research system I built at ResearchTech, with a 93.9% prompt-cache hit rate per firm.

Selected work

Systems I've built for clients. Research first.

Research and data pipelines

VC-startup matchmaking · Client under NDA

Startups stop pitching investors who were never a fit.

Read the case

Matches for Brightline · seed, B2B payments

  1. 01Northfield CapitalLeads seed rounds in B2B fintech. Two partners focus on payments. northfield.example
  2. 02Harbor Lane VenturesCheques fit the round. Backed a similar company in 2024. news.example
  3. 03Ostrava SeedInvests in CEE founders at this stage. ostrava.example
Fictional startup and funds.
Input
A startup describes itself to its own Claude, connected to the product over MCP.
What I built
A schema of what must be known about a fund to decide a match, and a multi-stage agent pipeline per fund that resolves conflicting sources and keeps a source on every claim. A reranking stage that updates the matches from the startup's own inputs, updates and call transcripts. Plus the product around it: backend, a UI to browse every VC, and auth.
Output
A ranked list of the VCs that actually fit, grounded in deep research on each fund. In use by the client.
Stablecoin evaluation · Credora

Credora enters a stablecoin name. Their spreadsheet fills itself.

Read the case

Stablecoin · USDT

Example USDT row in Credora's schema
License Type 1Digital asset service providerregulator.example/registerHigh
CategoryLiquid stables (blue-chip)framework.exampleHigh
Market Cap$150Bmarket-data.exampleHigh
Category Market Cap$240Bmarket-data.exampleMedium
Total Liabilities$148Bissuer.example/attestationHigh
Example values, not current data. One row per field in Credora's schema.
Input
Credora's schema and a stablecoin name.
What I built
A research pipeline with crypto data integrations and automated research that fills every field.
Output
The spreadsheet their team used to fill by hand, filled with a source and a confidence level on every value.
Agentic non-determinism finder · Hacken / QAN

400+ non-deterministic spots found in a blockchain engine.

Read the case

Hypothesis H-17 · map iteration in state hashing

run 01state root 0x8f3a…c21

run 02state root 0x8f3a…c21

run 03state root 0x41c9…7e0differs

run 04state root 0x8f3a…c21

Non-deterministic. Confirmed and reported.

Example log.
Input
A blockchain engine that has to behave the same way on every run.
What I built
An agentic system that generates hypotheses about where the engine could behave non-deterministically, then generates and runs many tests per hypothesis.
Output
400+ non-deterministic spots found and confirmed.
Open-sourced by Hacken on GitHub
Other AI systems

Email handling, chatbots, testing agents, video analysis and more.

Show all
  • Email handling in a CRM~2,000 emails a day sorted into orders, disputes and issues. About three people's work, with support replies in seconds instead of hours.
  • Healthcare chatbotPatients book a visit in one conversation. Hands over to a real person when needed.
  • Email order intakeFor a metal-products manufacturer: AI maps free-text customer emails to internal products and drafts replies. A 2-hour daily job dropped to ~2 minutes.
  • AI that tests softwareDrives UIs, fills and validates forms by itself.
  • Game feedback from video at scaleAnalyses every frame of gameplay videos on YouTube and Twitch, with scenes, player emotions and viewer comments, to collect frustration, bugs and other early feedback for game developers automatically.
  • Voice and video analysisMy oldest specialty, back to an MSc thesis on neural-network object tracking in video.
  • Invoice extractionDocument structuring and invoice data extraction.
  • And many more under NDA.Ask on a call.

Working together

You tell me what you need to know. I build the pipeline for it.

  1. 01Free

    Discovery call

    You bring
    What you need to know, the shape you need it in, and a sample list.
    You get
    A short workshop on goals, non-goals and the expected outcome, then a schema and a plan.
  2. 02Week 1

    Pilot

    You bring
    What you use today, as the baseline.
    You get
    A simple first version to test on your own sample, measured against that baseline.
  3. 03Week 2

    Working version

    You bring
    Feedback from the pilot.
    You get
    A good working version, built from that feedback.
  4. 04Ongoing

    Keeps improving

    You bring
    Feedback from real use.
    You get
    Continuous improvements, and recurring runs on a schedule or a signal if you need them.

Promise

If it doesn't beat what you use now, you don't pay.

I run the pilot on a sample of your own list. Then we compare it with how you do it today: your tool, your team or your vendor.

Cost
per entity
Coverage
fields you need, actually found
Quality
values that are right
Speed
to the full dataset
Pilot wins on at least one
You pay for the pilot
Pilot wins on none
You pay nothing

Two ways to buy

  • One-off dataset

    Your list, researched once to your schema.

  • Recurring datasets

    The same research, run again on a schedule or a signal, so the data stays current.

Delivered how you need it

API · MCP
Integrated with the tools and systems you already have.
HOSTED
I host the data and show it on a platform I set up for you.

Questions before a call

Is the discovery call free?

Yes.

How soon will I see results?

A simple pilot to test within one week, and a good working version in two. After that I keep improving it from your feedback.

Who owns the data?

The output data belongs to you.

What about paid data sources?

They run on your own keys and contracts, so data rights stay clean.

What does it cost?

It depends on the scope and how often the research runs. Pricing on a call.

What if the pilot doesn't beat what I use now?

Then you don't pay for it. The pilot has to be cheaper at the same quality, much better, or much faster than what you use today.

Book a free discovery call

Also available

Need something other than research?

AI systems, built end to end

Agents, automations, chatbots and AI products, with the backend, UI and infrastructure around them. See the work above.

Ask about a build

Workshops and team training

Workshops and training for engineering teams on AI-assisted development: setting up a harness well, what matters in practice, and choosing the right models. Past clients include Syndicode.

Ask about training
Illustrated portrait of Yuriy Babyak in a white T-shirt, leaning on a chair in front of a diagram, with cobalt and ochre geometric shapes.
Yuriy BabyakAI systems engineer · Warsaw / Kraków

About

I build the whole system, from the agent fleet to the UI. Depth is the point, and I make it cheap.

I start from the business need and the budget, ship a working version within a week, and improve it from real feedback.

About ten years in software and in AI since 2022, with an AI master's degree from before the LLM wave. Before going independent I was founding engineer at ResearchTech, where I built several research systems. The largest and most complex was the VC research system behind the numbers on this page.

Most AI-dev problems aren't model problems, they're harness problems.

From my talk in Kraków, Roo Code Poland / ActiveCampaign

Track record

LittleBit.devOwn studio
Research as a service, AI systems built end to end, training and consulting. Clients include Credora, Hacken and Syndicode.
ResearchTechFounding engineer
Built the VC research system and a ~30-step startup due-diligence pipeline.
WaypassFounder
AI platform for immigration lawyers handling Polish residence-permit cases.
GoelettAI and lead frontend
Customer-support automation and an email pipeline handling ~2,000 messages a day, about three people's work.
Hexagon GeospatialFrontend
Real-time event-monitoring system used at JFK Airport.

Hackathons

  • Winner, Google DeepMind x AI Tinkerers, WarsawA working MVP in 6 hours that analyses player videos to automate game playtesting.
  • Warsaw Model Trainers HackathonA harness of cooperating agents ran 150+ benchmarks and built the solution itself. A 4B model scored 46/60 on the Polish history matura, offline, on the size-to-quality Pareto line.
  • OpenAI hackathon with ElevenLabs24 hours.
  • Winner, ML hackathon 2019Fastest and most accurate sword tracking in video.

LinkedInGitHubyuriy.babyak@outlook.com

Bring one research question.

Book a free discovery callor email yuriy.babyak@outlook.com