Case study
Building an AI-Powered Pitch-to-Investor Matching Platform for a UK EdTech Startup
Plexteq partnered with FounderTribes, a UK entrepreneurial-education startup, to engineer the backend of a next-generation platform where founders record video pitches, receive structured mentor feedback, and are matched to the right investors through an AI/ML matching engine, delivered with an advanced RBAC security subsystem to protect founders' proprietary ideas, an integration test suite covering 100% of endpoints, and a fully containerised CI/CD pipeline with 3-minute deployments.
Project Highlights
Industry
EdTech / E-Learning (Entrepreneurial Education)
Market
UK / EU
Cooperation
2020 – 2021
Expertise
Backend Engineering, AI/ML Matching, RBAC & Data Security, Video Streaming, Test Automation, DevOps & CI/CD
Technologies
Node.js, Express, MongoDB, Redis, Docker, CircleCI, Brightcove, GCP, Firebase, InfluxDB, Agenda, Postman, Swagger, Zabbix, Grafana
Business Challenge
FounderTribes was built around a simple but demanding idea: give early-stage entrepreneurs a "learn, connect, raise" loop in which they pitch their business to a community of peers and investors, get their pitch audited to surface weak spots, work through targeted learning modules and one-to-one mentor sessions to fix those weak spots, then pitch again - and, once the business is ready, get introduced to investors who can fund it.
Delivering that loop as a product is far harder than it looks. Pitches are video-first, so the platform has to ingest, store, transcode, and stream founder video reliably at scale. Feedback is structured - mentors and reviewers score each pitch across weighted evaluation categories to produce an investability signal, so the data model has to capture nuanced, category-level review data rather than free-text comments. And the payoff of the whole loop, the investor introduction, depends on matching quality: connecting the right investor or mentor to the right founder means reconciling unstructured pitch narrative with hard structured constraints such as funding stage, geography, and industry, and doing it fast enough to feel real-time.
On top of the product complexity sat a security problem that, on this platform, was existential rather than incidental. Startups live and die by their unique know-how - the ideas, market insights, and unfair advantages inside a pitch are exactly the assets a founder cannot afford to leak. Yet the platform deliberately brings founders, investors, mentors/reviewers, and administrators into the same data space, each entitled to a very different slice of every record. Protecting founders' confidential material as it flows through that shared space - ensuring an investor, a reviewer, or a curious peer only ever sees what they are entitled to see, was treated as a top-priority requirement from the outset, which called for an access-control system far more granular and trustworthy than conventional role-based permissions.
Finally, FounderTribes was a small, fast-moving team that needed to iterate on the product continuously without a large manual QA function slowing every release, and without deployments becoming a source of risk. They engaged Plexteq to design and build the platform's backend end-to-end - with correctness, access control, and delivery speed treated as first-class requirements rather than afterthoughts.
Key Challenges
#1. Protecting Founders' Proprietary Ideas
The most sensitive assets on the platform are the founders' own the know-how and ideas inside every pitch. With founders, investors, mentors/reviewers, and administrators all sharing the same records but entitled to radically different views of them, a single over-permissive rule could expose confidential material to the wrong party. Safeguarding that material demanded an access-control subsystem more granular and more trustworthy than standard role-based permissions - one able to govern access down to the individual attribute, and to guarantee that a given role, or in some cases only the record's owner, could ever see or change it.
#2. Matching Unstructured Pitches to Structured Constraints
Good matches depend on combining two very different kinds of signal: the semantic content of a founder's pitch and hard structured parameters such as stage, geography, and prohibited industries. The engine had to rule out illegal pairings outright, score the remainder on a unified notion of "fit," and return the best candidates fast enough for interactive use.
#3. Video at the Core of the Product
The entire experience is anchored on founder video pitches, which meant reliable upload, transcoding, storage, and adaptive streaming were non-negotiable - without diverting the team into building and operating video infrastructure from scratch.
#4. Shipping Fast Without a QA Bottleneck
A small team iterating continuously could not afford slow, manual regression testing or fragile deployments. Quality and release velocity had to be engineered in, so that changes could reach production frequently and safely across multiple environments.
Solution Delivered
↳ Backend architecture on Node.js and MongoDB
Plexteq built the backend on Node.js and Express with MongoDB, split into two independently deployable services: a stateless API server for client traffic and a job server for scheduled and asynchronous work - media hand-offs, notifications, maintenance, via a MongoDB-backed scheduler.
Redis handled caching and rate limiting, and versioned migrations kept database changes safe. The domain model covered the whole "learn, connect, raise" loop: founders and business ideas, video pitches, weighted reviewer scoring, investability signals, communities, connections, and notifications.
↳ Advanced RBAC security subsystem
Because protecting founders' proprietary ideas was a first-priority requirement, Plexteq built a dedicated RBAC subsystem rather than relying on endpoint- or document-level permissions. Its defining property is granularity: access is enforced per role, per record, and per individual attribute - so the system can express rules like "an investor sees a founder's traction metrics but not their strategy notes," or "only the owner may read this field." That is what makes the confidentiality guarantee credible: sensitive parts of a record stay invisible to every other party sharing it.
The subsystem is model-driven - each role's rights are declared on the data model, next to the data they govern - and enforced automatically in both directions: an inbound guard strips attributes the caller cannot write, an outbound guard strips those they cannot read. It resolves nested and referenced objects recursively and supports an owner-only scope, so every new field's policy is defined once and enforced everywhere by default.
↳ AI/ML pitch-to-investor matching engine
At the heart of the "connect" step, a multi-stage pipeline turns pitches and investor profiles into a ranked, real-time shortlist:
-
Hard filtering. Boolean constraints and vector pre-filtering drop illegal pairs up front - geography exclusions, stage mismatch, prohibited industries — so scoring only sees viable candidates.
-
Feature encodings. LLM embeddings turn pitch text into dense semantic vectors, and feature scalers standardise the numeric parameters, giving every startup and investor a comparable representation.
-
Metric calculation. A two-tower embedding model with a weighted distance function blends pitch similarity and parameter similarity into one fit score
-
Ranking & search. An approximate-nearest-neighbour index (HNSW / ScaNN) returns the K nearest investors for a startup in real time.
The same principle - semantic content plus structured constraints also powers in-platform content recommendation.
↳ Video pipeline
Rather than build video infrastructure, Plexteq integrated a managed platform (Brightcove) for pitch ingestion, transcoding, and adaptive streaming, with Google Cloud Storage for supporting assets such as thumbnails, avatars, and pitch decks. Uploads ran through a streaming multipart layer, and per-video view-marks fed watch progress and engagement back into the product.
↳ Integration test suite and QA automation
Plexteq built a comprehensive integration suite against an ephemeral in-memory MongoDB that exercises the API end-to-end rather than mocking it, with a dedicated suite per domain area and. It reached 100% endpoint coverage and roughly 95% of business logic, tracked on every run - turning regression testing into an automated gate and cutting manual QA to the bare minimum.
↳ Containerisation and CI/CD
The system was containerised with Docker and deployed on Kubernetes. A CircleCI pipeline builds, tests, and ships changes automatically across three environments - development, staging, and production at an average of about three minutes per deployment, making releases routine rather than risky.
↳ Real-time analytics for founders
Plexteq also turned the platform's telemetry into a product feature. Usage and engagement metrics stream in real time to InfluxDB and surface through Grafana dashboards built for founders, so each founder can watch their advertising and marketing campaigns and overall pitch activity as it happens - a live feedback loop that complements structured mentor scoring with hard behavioural data.
↳ Extending to university entrepreneurship programs
As the platform matured, Plexteq extended it beyond individual founders to host entrepreneurship programs inside universities. Institutions could join as first-class participants, bring students in as entrepreneurs, and connect them with real investors for pitch assessment and feedback. It reused the existing building blocks - pitch-and-review workflow, investability scoring, matching engine, and RBAC subsystem adding only the organisational layer to run student cohorts within an institution.
Key Results
100%
of API endpoints covered by automated integration tests
~3 min
average deployment time across a containerised CI/CD pipeline
Real-time
investor matching returned interactively via the AI/ML pipeline, not in batch
Business Outcome
Plexteq delivered the FounderTribes backend end-to-end - data model, RBAC security subsystem, AI/ML matching engine, video pipeline, test suite, and deployment pipeline as one coherent system. The RBAC subsystem met the platform's top priority directly: attribute-level access, enforced automatically on every read and write, keeps founders' proprietary ideas protected in a space shared with investors, reviewers, and peers, so each party sees only what it is entitled to.
Automated testing and a containerised CI/CD pipeline let a small team ship continuously - regressions caught before release, deployments routine rather than risky. Founders gained a real-time feedback loop on campaign and pitch performance through InfluxDB and Grafana, and the architecture proved extensible: the same core was later expanded to run university entrepreneurship programs without rebuilding the pitch, review, matching, or access-control foundations.
Capabilities delivered through the engagement::
-
An advanced RBAC security subsystem protecting founders' proprietary ideas, with attribute-level access enforced on both read and write paths
-
A real-time AI/ML matching engine connecting founders to best-fit investors and mentors
-
A video-first pitch experience on managed streaming infrastructure, plus real-time founder analytics via InfluxDB and Grafana
-
Architecture extended to university entrepreneurship programs on the same core platform
Plexteq proved itself a dependable product engineering partner in delivering FounderTribes' core mission: a platform where founders sharpen their pitch and connect with the right investors and mentors. We built that experience end-to-end and made delivery fast and dependable, so the product could keep evolving with confidence.
Just as importantly, we treated the protection of founders' proprietary ideas as a first priority, and delivered a foundation the client could build on - one flexible enough to grow from individual founders into university entrepreneurship programs, and to support their ongoing product and training ambitions.