Skip to content
IdeaScore
IdeaScore AI

A co-evaluator that does its homework and shows its sources

IdeaScore AI reads each application and its deck, researches the claim on the live web, and scores it against your rubric with a citation under every criterion. Its score sits beside the human scores and never replaces them.

Book a demoSee pricing
IdeaScore AI
Deep research
46.5 / 50 (93%)

Halcyon Grid · #0142 · CleanTech · TRL 6

  • Innovation and novelty9.5 / 10
  • Customer desirability9 / 10
  • Team and execution9 / 10
  • Technical feasibility9.5 / 10
  • Business viability9.5 / 10

“Grid-edge storage for rural feeders is under-served; two comparable ventures found, neither targets sub-1 MW feeders. The founder's prior work at a state utility is verifiable.”

Sources (6)

Market sizing report, 2025Competitor scan, 4 companiesPatent search, IN and PCTFounder profileState EV and storage policyPitch deck, slide 7

IdeaScore AI may make mistakes. Use it alongside human judgement.

What it researches

Six questions asked of every application

The same checks a diligent reviewer would run, on every submission rather than on the ten they had time for.

  • Market size

    How large the stated market is, and whether the applicant's own number holds up.

  • Competitors

    Who else is doing this, how far along they are, and what the applicant is not mentioning.

  • Intellectual property and prior art

    Patent and publication searches in the relevant jurisdictions for the core claim.

  • Mentor and founder credibility

    Whether the stated affiliations, prior work and advisers can be verified in public records.

  • Policy alignment

    Whether the venture fits the schemes, missions and state policy the programme cares about.

  • Technical feasibility

    Whether the approach is plausible at the stage of development claimed.

How a research run works

From application to cited score

A run takes a few minutes and produces a score, a rationale and a source list a reviewer can check.

  1. Reads the application and deck

    Every answer, plus the uploaded pitch deck, with claims and numbers pulled out slide by slide.

    Halcyon Grid · #0142

    90 fields · deck, 14 slides

  2. Searches the live web

    Targeted queries per criterion against current sources, not a snapshot from training.

    Queries run

    market · competitors · patents

  3. Scores each criterion with citations

    Your rubric, your weights, with the sources that support each band recorded against it.

    Scored

    46.5 / 50

  4. Writes a reviewer summary

    A short rationale in plain language, so an evaluator starts from evidence rather than a blank page.

    Sources

    6 cited

How IdeaScore AI fits into a call

Evidence beside judgement

The AI score never overwrites a human one

Both scores sit on the same row against the same rubric. Where they agree, a reviewer moves on quickly. Where they disagree, the gap is the interesting part: the AI found a competitor the panel had not, or the panel met the founder and the AI did not. The blended view is a lens on the table, not a number written into the record — the human score remains the one of record throughout.

  • Same rubric, same criteria, same total
  • Disagreement shown rather than averaged away
  • Blended is a view, never the stored score
  • Every AI criterion links to the source behind it
Seed Fund 2026 · Cohort III
Halcyon Grid#0142

CleanTech · TRL 6 · same rubric, out of 50

Human panel

46.5

3 evaluators, mean

Met the founder at the campus clinic. Strong on execution, cautious on the market claim.

IdeaScore AI

44

6 sources cited

Found two comparable ventures the panel had not seen; neither targets sub-1 MW feeders.

Blended view

A lens on the table. The human score stays the score of record.

45.25 / 50

Pre-screening

Catch the broken ones before a human opens them

Rule-based checks and IdeaScore AI run over the pile as it arrives: mandatory answers missing, a deck that will not open, a budget that does not add up, a venture that has already applied under another name. Flagged records come with a reason, and a programme admin decides whether to ask the applicant to fix it or to set it aside. Nothing is rejected automatically.

  • Completeness checks on required answers and files
  • Duplicate and near-duplicate detection across the call
  • A reason on every flag, not just a marker
  • The decision stays with a person
Seed Fund 2026 · Cohort III

Pre-screen results

5 flagged of 181

  • #0166
    Northwind CellsIncomplete

    Budget table missing from stage 3

  • #0158
    FarmlightFile

    Deck will not open — corrupt PDF

  • #0151
    Quorum BioDuplicate

    Matches #0119 on venture and contact

  • #0147
    Sable OpticsClaim

    Stated market figure not supported by any source found

  • #0143
    KelplineIncomplete

    Two required answers left blank

Flags are suggestions. Nothing is rejected without a person deciding.

Control and cost

You set how much of it runs

IdeaScore AI is metered by evaluations per day, set by your plan and visible as a meter your team can see before they start a batch. A request log records every run with the submission, the person who started it and the outcome. When an application is edited after an evaluation, the evaluation is flagged as stale so nobody reads an old answer, and any programme can turn the module off entirely.

  • A daily quota with a visible meter
  • A request log with actor, submission and outcome
  • Staleness flags when the application changed after the run
  • Off per programme, with one setting
Seed Fund 2026 · Cohort III

Daily quota

14 of 20 used

Resets at midnight. 70% of today’s allowance on the Professional plan.

Request log

Today

  • #0142
    Completed

    Programme admin

    16:52
  • #0139
    CompletedStale · application edited since

    Programme admin

    16:44
  • #0127
    Completed

    Ravi, programme staff

    15:58
  • #0118
    Completed

    Programme admin

    15:31
  • #0106
    Superseded

    Programme admin

    11:02

IdeaScore AI may make mistakes. Use it alongside human judgement.

In the app, beside the work

IdeaScore AI is a page in the programme, not a separate tool. A reviewer opens a submission, reads the rationale, follows a source and scores it themselves.

IdeaScore AI may make mistakes. Use it alongside human judgement.

A real IdeaScore AI evaluation card in dark mode, with per-criterion scores, an overall assessment and the web research behind it.

A real evaluation from the app, run on a synthetic test submission.

Specification

Everything in IdeaScore AI

  • Grounded web research

    Live searches per criterion, with sources recorded on the evaluation.

  • Cited criteria

    Every score band carries the source that supports it.

  • Pitch-deck extraction

    Claims, numbers and the ask pulled out of the uploaded deck.

  • Reviewer summary

    A plain-language rationale an evaluator can start from.

  • Pre-screening

    Completeness and duplicate checks before a human opens the pile.

  • Cohort benchmarking

    How a submission sits against the rest of the same call.

  • Daily quota

    Evaluations per day by plan, with a meter before you start a batch.

  • Request log

    Every run recorded with actor, submission and outcome.

  • Staleness flags

    Evaluations marked stale when the application changes underneath them.

  • Per-programme switch

    Turn the module off for a programme that should be human-only.

  • Rubric-aligned

    Scores against your criteria and weights, not a generic scale.

  • Human score of record

    The AI result is always an additional column, never a replacement.

What people ask before switching it on

Does IdeaScore AI replace our evaluators?
No. It produces an additional column beside the human scores. The human score stays the score of record, promotion decisions are made by people, and a programme can run with the module switched off entirely.
What does it not do?
It does not approve, reject or promote anything. It does not read applicant data from outside the submission, it does not score criteria you have not defined, and it does not claim certainty — every evaluation carries the reminder that it may be wrong.
Where do the sources come from?
From live web searches run at evaluation time, plus the application and the uploaded deck. Each source is recorded on the evaluation so a reviewer can open it and check the claim themselves.
How is the quota counted?
One evaluation of one submission is one unit against your daily allowance. Re-running an evaluation after the application changed counts again. Pre-screening checks do not.
What data leaves our server?
The text of the application and the extracted deck content needed for the research, sent to the evaluation service at run time. Nothing is retained for training. On a self-hosted deployment you control the outbound configuration, and the module can be left off.
Can it be turned off for one programme?
Yes. The module is a per-programme setting as well as a plan feature, so a sensitive call can run human-only while another programme in the same organisation uses it.

See IdeaScore run a call with your own rubric

A 30-minute walkthrough with a founder, using your programme's form and criteria.

Book a demoStart free