DOUGHDILLY

Venture capital has an information problem.

Investors make decisions about companies whose futures are uncertain, often relying on pitch decks, financial projections, founder interviews, and personal networks. But startups are constantly changing, and the evidence available during an initial meeting can become outdated within weeks.

Consider two startups.

The first reports $2 million in annual recurring revenue, impressive gross margins, and rapid hiring. But further investigation reveals that a significant portion of revenue comes from temporary pilots, margins exclude AI inference costs, and engineering activity has slowed.

The second reports more modest financial results but demonstrates improving customer retention, increasing product velocity, strong customer engagement, and founders who consistently address earlier concerns.

Which company is actually becoming the better investment?

That question inspired DoughDilly.

Our research into startup evaluation in the AI era revealed that traditional venture metrics are increasingly insufficient on their own. AI-native companies can scale with unusually small teams, generate revenue that is not necessarily durable, and face infrastructure costs that traditional SaaS benchmarks fail to capture.

We also identified a category of signals that conventional pitch-deck analysis often overlooks: the operational evidence companies leave behind.

GitHub activity, hiring patterns, product releases, pricing changes, customer retention, founder communication, and responses to previous investor objections can reveal meaningful changes before they become visible in headline financial results.

Another important insight was that venture investors frequently encounter the same startup more than once. A company rejected for insufficient traction might return months later with paying customers and a stronger product. Yet the original reasons for passing may remain buried in emails, meeting notes, or individual investors' memories.

We wanted to build a system that could preserve that context and help investors evaluate progress rather than repeatedly starting from scratch.

We also believed that AI-powered investment analysis should be transparent and evidence-aware. A model should not invent confidence when information is missing, and founders should not be unfairly disadvantaged because they lack prestigious credentials or established investor networks.

DoughDilly emerged from the intersection of three ideas: longitudinal evaluation, evidence-based AI reasoning, and accountable investment decision-making.

What it does

DoughDilly is an AI-assisted startup and founder evaluation platform designed for venture capital firms, angel investors, accelerators, and investment analysts.

It combines historical startup performance, structured scoring, AI-powered progress analysis, competitive intelligence, and interactive workflow visualizations into a single application.

Time-sequenced startup evaluation

Unlike traditional scoring tools that generate one assessment, DoughDilly lets investors examine how a company's evaluation changes across multiple meetings.

Users can navigate historical snapshots to compare annual recurring revenue, net revenue retention, burn multiple, gross margin, customer retention, team size, investment scores, red flags, and investor commentary.

The evaluation timeline shows how new evidence changes the understanding of a startup's performance.

This helps investors identify whether a company is improving, deteriorating, or revealing previously unknown risks.

Evidence-based investment report cards

DoughDilly evaluates startups across ten weighted dimensions:

Team

Founder-market fit

Market and timing

Product

Traction

Retention and durability

Unit economics

Defensibility

Founder learning rate

Deal terms and risk

The application calculates a Merit score using deterministic TypeScript logic and displays a separate Fund Fit score.

This distinction is essential: a company can be an excellent business without matching a particular fund's investment strategy.

DoughDilly also includes an abstention mechanism. When fewer than six of the ten evaluation dimensions contain evidence, the system displays "Not enough data" instead of generating a misleading Merit score.

The result is a scoring framework designed to acknowledge uncertainty rather than hide it.

AI-powered founder progress assessments

Investors can submit founder updates, investor letters, or meeting notes for AI-assisted analysis.

Using Neecroi Ninoo, DoughDilly compares the new information against the startup's existing KPIs and generates a structured assessment.

The output includes:

A summary of the founder's progress

Expected directional changes in relevant KPIs

Potential inconsistencies and red flags

Confidence levels

Targeted follow-up questions

An evaluation note suitable for an investor's internal records

The model is instructed not to invent financial numbers and to mark changes as unclear when the available evidence is insufficient.

This allows investors to turn unstructured founder communications into more actionable diligence questions.

Competitive intelligence powered by Akta

DoughDilly incorporates real-world company-enrichment data retrieved through Akta's API.

We used Akta to obtain firmographic, management, financial-estimate, digital-presence, strategic, and trust signals for companies representing different startup categories.

These records inform our fictional startup evaluation scenarios and provide competitive context for market and defensibility assessments.

For example, a fictional accounts-payable startup may claim a substantial proprietary data advantage. Comparing that claim with the publicly reported scale of an established competitor helps illustrate whether the claimed advantage is genuinely defensible.

The demonstration uses fictional applicants while grounding parts of their competitive context in real enrichment data.

Visual agentic evaluation workflow

DoughDilly includes an interactive evaluation pipeline that illustrates the stages of startup due diligence.

The workflow represents specialized responsibilities for intake, external research, founder evaluation, financial metrics, historical context, verification, scoring, and auditing.

Users can explore the process through an agent graph, a holographic 3D visualization, and an activity stream showing the sequence of evaluation steps.

The current hackathon demonstration uses scripted agent events to replay the workflow consistently, while the founder progress-assessment feature performs actual AI inference.

This combination makes the proposed multi-agent architecture understandable without requiring judges to interpret raw backend logs.

AI traceability through BlockConvey PRISM

DoughDilly integrates BlockConvey PRISM to record and retrieve AI assessment traces.

After a successful model assessment, the server-side integration can submit the prompt, structured output, model identifier, execution latency, and session metadata to PRISM.

We also built a dedicated PRISM traces dashboard that retrieves assessment records and organizes them by startup.

This allows users to inspect the inputs and outputs behind AI-generated conclusions, supporting transparency and debugging.

The integration provides a foundation for expanding into systematic fairness, consistency, and explainability evaluations.

Agent activity recording through Roundtable

DoughDilly also integrates Roundtable's API to record completed AI assessment activities.

Each record can include the evaluated startup, model, confidence level, KPI impacts, identified red flags, execution latency, and assessment summary.

This adds an activity-tracking layer alongside the PRISM traces and the investor-facing evaluation results.

Interactive demo scenarios

The application includes multiple fictional startup scenarios representing different diligence outcomes.

Some companies demonstrate improving traction and credible execution, while others reveal contradictory financial claims, weakening operational signals, or insufficient evidence.

Investors can move between applicants, explore historical snapshots, replay the evaluation workflow, and submit founder progress updates.

Together, these features demonstrate how DoughDilly helps investors move beyond static startup assessments toward a more continuous, evidence-driven evaluation process.

How we built it

We began by researching how startup and founder evaluation is changing in the AI era.

Our research examined traditional venture metrics, alternative company signals, founder behavior, stage-specific evaluation criteria, and the challenges of evaluating AI-native businesses.

We then translated these findings into a technical design centered on ten evaluation dimensions, historical scoring, specialized AI responsibilities, and explainable decision support.

Application architecture

We built DoughDilly using React, TypeScript, TanStack Start, Vite, Tailwind CSS, and reusable UI components.

Lovable helped accelerate the frontend development process, while GitHub served as the source-code repository.

The application includes historical evaluation timelines, KPI visualizations, weighted scorecards, founder progress analysis, agent graphs, and an interactive holographic workflow representation.

Deterministic scoring engine

We implemented a TypeScript scoring engine that calculates startup Merit across ten weighted dimensions.

Each dimension can receive a score from one to five or remain unknown when evidence is unavailable.

The engine normalizes available scores, applies their configured weights, and incorporates a bounded verification adjustment.

When fewer than six dimensions contain evidence, the scoring function abstains rather than returning a number.

This separates numerical calculations from language-model generation and makes the scoring formula inspectable and testable.

Neecroi Ninoo integration

We integrated Neecroi's OpenAI-compatible model gateway using the Vercel AI SDK.

The application is configured to use Claude Sonnet 5.5 for founder progress assessments.

A server-side function receives the startup context, existing KPIs, and founder update, then submits the information to the model.

The response follows a structured JSON contract containing KPI impacts, red flags, follow-up questions, confidence, and an evaluation note.

Zod validates the response before it is returned to the user.

Akta integration

We retrieved real company-enrichment responses from Akta for five companies representing different startup categories.

The enrichment data was incorporated into synthetic evaluation scenarios as competitive and market context.

This allowed us to demonstrate how external company intelligence can help challenge or substantiate claims about traction, scale, and defensibility.

BlockConvey PRISM integration

We implemented server-side PRISM API calls for trace ingestion and retrieval.

Assessment traces include model metadata, inputs, outputs, latency, and identifiers that group related evaluations.

A dedicated dashboard retrieves these records and presents them in an inspectable format.

Roundtable integration

We connected DoughDilly to Roundtable's activity-recording API.

After an AI assessment, the application can submit a structured activity summary containing the startup, model, confidence, KPI impacts, red flags, and latency.

Agentic workflow visualization

We developed a staged evaluation experience representing specialized agents and their interactions.

The frontend includes an animated agent graph, a holographic 3D visualization, and a process flowchart.

Scripted evaluation events power the current demo playback, allowing the system to demonstrate complex evaluation paths consistently.

Technologies used

Frontend: React, TypeScript, TanStack Start, Vite, Tailwind CSS, Recharts, Lovable

AI: Neecroi Ninoo, Claude Sonnet 5.5, Vercel AI SDK, Zod

Data and intelligence: Akta API, synthetic startup evaluation datasets

Observability and activity: BlockConvey PRISM, Roundtable

Development: GitHub, TypeScript, Vitest

Challenges we ran into

Evaluating startups as evolving businesses

Most evaluation tools assume a startup can be represented by one set of metrics.

We wanted to preserve historical context and show how new information changes an investment assessment.

This required designing evaluation snapshots, time-dependent KPI displays, historical investor notes, and consistent scoring across meetings.

Avoiding misleading AI-generated confidence

One of the biggest challenges was preventing the platform from assigning precise scores when supporting information was missing.

We addressed this by implementing an abstention gate and allowing individual evaluation dimensions to remain unknown.

This reflects our belief that an AI system should be able to say it does not have enough information.

Distinguishing real traction from headline metrics

AI-native startups create unusual diligence challenges.

Revenue may come from temporary pilots, infrastructure expenses can distort gross margins, and product demonstrations may not reflect real production performance.

We designed the evaluation framework and synthetic scenarios to make these issues visible rather than accepting reported metrics at face value.

Separating company Merit from Fund Fit

A startup can be fundamentally strong but unsuitable for a particular investor.

We needed to avoid combining these judgments into one misleading score.

Displaying Merit and Fund Fit independently allows investors to distinguish business quality from strategic alignment.

Integrating multiple external services

Connecting Neecroi, Akta, PRISM, and Roundtable required handling different authentication methods, request formats, response structures, and failure conditions.

We used server-side integrations to protect credentials and structured validation to make model responses predictable.

Making an agentic system understandable

Complex AI workflows can be difficult for users to interpret.

We wanted the evaluation process to be visible rather than hidden behind a single button.

Building the animated workflow, holographic visualization, and synchronized event display helped translate a technical architecture into an accessible user experience.

Accomplishments that we're proud of

Our biggest accomplishment is building a research-grounded startup evaluation experience that prioritizes a company's trajectory rather than a single pitch-deck snapshot.

We implemented a ten-dimensional scoring engine with an evidence threshold that allows the system to abstain when information is insufficient.

We separated Merit from Fund Fit, preserving an important distinction that traditional one-number scoring systems often overlook.

We developed a working AI-powered founder progress-assessment feature that interprets updates, identifies potential inconsistencies, and generates actionable diligence questions.

We integrated four external technologies into the application and its demonstration data: Neecroi Ninoo for AI reasoning, Akta for competitive enrichment, PRISM for traceability, and Roundtable for assessment activity recording.

We also built interactive historical timelines, KPI visualizations, investment report cards, agent graphs, and a holographic evaluation experience.

Another accomplishment was creating realistic synthetic startup scenarios informed by real company-enrichment data, allowing us to demonstrate both promising and concerning investment trajectories.

Most importantly, we designed DoughDilly around a principle that matters to us:

AI should improve the quality and accountability of investment decisions, not replace human judgment with an unexplained algorithm.

What we learned

Building DoughDilly reinforced that startup evaluation is fundamentally a problem of understanding change.

A company's current metrics matter, but the direction of those metrics, the credibility of their underlying evidence, and the founders' responses to new information can be equally important.

We learned that AI-native businesses require more careful interpretation of traditional financial indicators.

Durable revenue, inference-adjusted margins, production reliability, and efficiency per employee can reveal information that headline ARR or headcount figures conceal.

We also learned that founder behavior should be evaluated through observable evidence rather than relying excessively on credentials, presentation quality, or professional networks.

A founder who acknowledges uncertainty, addresses objections, and improves execution between meetings may demonstrate qualities that are not captured in a static application.

Another important lesson was that an AI system that refuses to guess can be more useful than one that always produces an answer.

The abstention gate became one of our most important design decisions because it turns missing evidence into an opportunity for better questions.

We also learned the value of combining deterministic calculations with language-model reasoning.

LLMs can interpret narrative information and identify possible contradictions, while programmatic scoring helps keep numerical results consistent and inspectable.

Finally, working with PRISM and Roundtable reinforced the importance of traceability and activity records when AI is used in consequential decision-support systems.

Explainability should not be an afterthought. It should be part of how the application is designed, evaluated, and experienced.

What's next for DoughDilly

Our long-term vision is to turn DoughDilly into a continuous investment intelligence platform that helps investors evaluate companies throughout their development.

Persistent startup and fund memory

We plan to introduce persistent profiles, historical meeting records, investment decisions, and fund-specific evaluation context.

This would allow investors to revisit previously evaluated companies without losing the reasoning behind earlier assessments.

Fully automated multi-agent due diligence

We want to extend the current scripted evaluation workflow into an operational multi-agent system that can ingest pitch decks, meeting transcripts, financial documents, and independently sourced company signals.

Specialized agents would collect evidence, verify claims, evaluate founders, normalize metrics, and prepare structured investment assessments.

Founder learning-rate measurement

We plan to automate comparisons between successive founder meetings.

The system could identify which claims became more precise, which objections were addressed, which metrics changed without explanation, and which commitments were fulfilled.

Pass replay and anti-portfolio intelligence

We want DoughDilly to remember why a fund previously passed on a startup and reassess those concerns when the company returns.

Over time, this could help investors understand which assumptions were correct, which opportunities were overlooked, and where their investment process could improve.

Fund-specific evaluation

Future versions could support configurable scoring weights, sector-specific benchmarks, investment theses, and decision thresholds.

This would allow different funds to evaluate startups according to their own strategies while preserving a consistent underlying evidence framework.

Deeper fairness and explainability auditing

We plan to expand the PRISM integration beyond assessment tracing into systematic checks for evaluation consistency, potential bias, and explainability.

We are particularly interested in reducing the risk that founder identity, educational background, or access to established networks disproportionately influence investment assessments.

Continuous monitoring and alerts

Eventually, DoughDilly could monitor new company developments, detect meaningful changes in operational signals, and alert investors when an earlier investment assessment deserves reconsideration.

From due diligence to investment intelligence

Our ultimate goal is to help investors build a continuously improving understanding of the companies they encounter.

We envision a system that remembers previous decisions, identifies changing opportunities, surfaces overlooked founders, and makes the evidence behind investment recommendations easier to inspect.

DoughDilly is not about replacing venture capitalists. It is about giving them better memory, better evidence, and better questions.

Because the most important question isn't whether a startup looks impressive today.

It's whether the evidence shows that it is becoming an extraordinary business.