✍️Inside Our DDVC London… | Data Driven VC Summary

1. Key Themes

Theme 1: Sourcing 2.0 — building proprietary, rerunnable pipelines that beat the data providers

Fence in your inputs to avoid hallucinated research

One team built a "second brain" that reads only from trusted sources instead of open-ended LLM search.

"They'd seen too many open-ended LLM searches return unreliable or invented results. The prototype scanned 300+ news items, kept what scored above a threshold, tracked how often focus areas and companies were mentioned, wrote briefs that link back to their sources, and pushed anything interesting to Slack."

The trade-off is flexibility: "the thesis is built into the code, so a generalist fund would need more work to adapt it."

Pre-CRM sourcing: finding companies before the data providers do

The third-place team started from incubators, accelerators, spin-outs and grants, then expanded via LinkedIn signals.

"They scraped about 7,900 companies, then their LinkedIn pages, their people, and the posts those people wrote about other startups. That added roughly 7,000 to 8,000 more leads. From about 15,000 signals, the agent narrowed the pool to around 190 companies that fit their theses."

Never miss a round: automated weekly digests with routing and outreach

"It collects the past week's rounds in fintech and AI infrastructure from Specter, adds rounds Specter missed, enriches each company, and routes each one to the person covering that vertical... A second step finds founder emails, matches them to the right thesis, and drafts outreach for a human to approve."

Theme 2: Deal ownership and workflow plumbing matter more than model cleverness

Unowned deals die in shared inboxes

The winning team's sharpest observation was about accountability.

"Today, new deals land in a shared inbox, and as one member put it, nobody really looks at it. Once a deal has a named owner and a deadline, it gets handled."

Deal Radar: an operating system for deal flow

"Routing based on expertise, geography, language, current capacity and existing relationships... A personal queue for every investor, with priorities, deadlines and to-dos, plus an admin view of who is overloaded... Inbox triage that classifies everything coming in... Document parsing from intro emails straight into the CRM... A knowledge graph with an assistant you can query."

Large funds are bottlenecked by process, not tooling

"Investors inside larger funds pointed to a different bottleneck: approval processes that touch many internal functions and can take weeks or even months."

Theme 3: Benchmarking against fund history, with data-quality gates

Benchmarking agent for founder calls and diligence

"The output is a dashboard with KPI comparisons, flags showing which data is verified, a score whose weights reflect confidence in each metric, and, most usefully, a list of questions to ask next. They ran it on real fund data, which impressed the judges."

Two rules: compare like with like, and gate on data sufficiency

"Match benchmarks on business model, stage and time period. A SaaS company today shouldn't be benchmarked against one from five years ago." "Add a gate that checks whether there's enough data to benchmark at all. At seed, where team and market dominate, there often isn't."

Theme 4: The market gap — an enterprise-grade agent layer for private funds

No compliant agent layer exists yet

"There are plenty of sourcing platforms and portfolio management tools. There seems to be no compliant, enterprise-grade agent layer built for private funds. Every fund in the room is assembling its own from separate pieces."

Convergence toward a unified system for running a fund

"Deal Radar won because it pointed furthest towards where venture is heading: one connected system for running a fund."

2. Contrarian Perspectives

Delay AI scoring so it doesn't bias humans

Most would assume AI scoring should be applied as early as possible to save time. The Deal Radar team argued the opposite.

"Showing an AI score too early can bias people and make them stop doing the work. Their approach: humans assess first. Once a deal reaches the IC discussion stage, an LLM runs scenario models as a second opinion, and it does flag risks the team hadn't considered."

Another participant scores entirely by hand: "1 to 5 after meeting the CEO, because an LLM can't see what happened in that conversation." Consensus: "The earlier the stage, the more guesswork. The later the stage, the more a model can add."

Funds may not want one all-in-one system

Against the "single platform" narrative, the judges asked:

"Do funds really want everything in one place? Or do they want focused tools whose output is easy to check, with humans clearly in charge of judgment?"

"Build vs. buy" — AI's last 10-20% is the expensive part

"AI gets you 80 to 90% of a tool like this in a few hours. The last 10 to 20% might take another 30 hours. Before you start, ask what a finished, maintained version is actually worth to you, and whether you can buy it."

Don't ask an LLM to re-scrape; have it write the script

"Asking an LLM to scrape again next week returns a different format every time and costs more. A script gives the same structure every run, so you can see what actually changed."

3. Companies Identified

Company: HSBC

Description: Global bank; co-host of the London hackathon.
Why mentioned: Partner and venue sponsor of the DDVC Builders Hackathon.
Quotes: "Last week, we ran the first DDVC Builders Hackathon in London, co-hosted with our partner HSBC." / "A big thank you to HSBC for hosting us, keeping everyone fed, and sending every builder home with an awesome cap!"

Company: Specter

Description: Startup/funding-round data provider.
Why mentioned: Data source for the Monday round digest, and also shown to have coverage gaps.
Quotes: "It collects the past week's rounds in fintech and AI infrastructure from Specter, adds rounds Specter missed, enriches each company..."

Company: Claude (Anthropic)

Description: LLM used by participants.
Why mentioned: Used to generate reusable scrapers.
Quotes: "Claude wrote Python scrapers the team can rerun each week."

Company: Affinity

Description: AI-first private capital CRM; newsletter sponsor.
Why mentioned: Sponsor promoting a PE-origination webinar.
Quotes: "Brought to you by Affinity – The AI-first private capital CRM" / "With mid-market deals drawing 8 to 15 buyers and deal volume down 34% in H1 2026, a big network no longer gives PE firms an edge."

Company: Armira

Description: PE firm.
Why mentioned: Its Head of Origination is a webinar speaker (sponsored content).
Quotes: "Join Niklas Krusche, Head of Origination at Armira..."

Company: Limited Partner List (LimitedPartnerList.com)

Description: LP database product.
Why mentioned: Suggested as the "buy" alternative to a half-built LP-mapping tool.
Quotes: "If this sounds interesting, check out LimitedPartnerList.com ;-)"

Company: Goodwin & Harmonic

Description: Law firm and VC data/AI platform, respectively.
Why mentioned: Co-hosts of the DDVC x Slush Breakfast in Helsinki.
Quotes: "Join our DDVC x Slush Breakfast in Helsinki Nov 18 with Goodwin & Harmonic"

Company: DDVC Lab

Description: Data Driven VC's product/community offering.
Why mentioned: Hackathon prize.
Quotes: "...the Deal Radar team, who won an annual 'RUN' subscription to the DDVC Lab."

4. People Identified

Person: Andre Retterath

Description: Author and host of Data Driven VC.
Why mentioned: Organizer of the hackathon and newsletter author.
Quotes: "Hi, I'm Andre and welcome to my newsletter Data Driven VC which is all about becoming a better investor with data and AI."

Person: Niklas Krusche

Description: Head of Origination at Armira.
Why mentioned: Speaker at the sponsored Affinity webinar.
Quotes: "Join Niklas Krusche, Head of Origination at Armira, and Veronica Mora... on October 21 at 1 PM ET."

Person: Veronica Mora

Description: VP of Customer Success at Affinity.
Why mentioned: Co-speaker at the sponsored webinar.
Quotes: "...Veronica Mora, VP of Customer Success at Affinity, on October 21 at 1 PM ET to learn how to spot a dormant network..."

5. Operating Insights

Convert recurring AI tasks into scripts

For anything on a weekly cadence, have the LLM write code once rather than re-running prompts.

"Turn prompts into scripts for anything that runs every week to reduce costs."

Wire every output into the CRM and relationship graph

Lead lists only matter if they show who can warm-intro.

"Every result is checked against the CRM. Each company is flagged as new or already known, with any existing relationship shown so the team knows who can make a warm intro." Also: "A list of companies is only useful once it leads to a next step."

Separate sourcing from diligence, and clean data first

Early judgment shouldn't prune the raw pipeline, and agents are only as good as the underlying data.

"Sourcing is kept separate from diligence, so early judgment doesn't filter the raw pipeline." / "Clean your data before building agents on top of it. Several teams hit this wall within three hours."

6. Overlooked Insights

Model switching changes scoring behavior, and scoring methods differ by sector

"The team also noted that switching models changes how the scoring behaves, even with the same prompt." For consumer deals, scoring can use "web traffic growth, Instagram, Trustpilot and app store data. For deep tech, those numbers mostly don't exist, so the LLM has to reason from the thesis." This implies model-version changes can silently shift deal prioritization and need monitoring.

Dormant relationship networks are a sourcing problem, and deal volume is shrinking

From the sponsored section: "deal volume down 34% in H1 2026" and a focus on helping firms "make relationships belong to the firm instead of individuals" and "use introductions as a leading indicator of sourcing health." Though promotional, it signals that institutionalizing relationship capital is becoming a competitive differentiator in a thinner deal market.