Insights, Not Transcripts
Raw conversation data becomes tagged themes, emerging risks and opportunities, and the topics shaping the pipeline – refreshed on a schedule, not dug out by hand.
Bridj
Tech Lead Consulting · Embedded since 2025
Embedded tech lead since August 2025 for a UK conversational-and-operational-intelligence platform – the layer that connects what customers actually say to what a business does next. Architecture, hands-on engineering across a 17-service system, and the final sign-off on every release.
- Embedded as tech lead since
- 08/2025
- Services across the platform
- 17
- Data-source integrations
- 7
- Production AI agents
- 4
Intelligence for the Whole Business
Bridj turns conversations, calls and customer interactions into one source of truth – surfacing what would otherwise slip through.
Ask Bridj
A conversational AI layer that answers questions over an organisation's own insights and data, drafts content from what's resonating, and writes executive briefings – in natural language.
Content That Lands
Campaign and content generation grounded in what customers actually said – built from real conversations rather than guesswork, and refined interactively.
What We Built Underneath
The product is one screen. Behind it, a pipeline carries messy, multi-source conversation data all the way to answers.
Connect & Sync
OAuth connectors pull data from seven external systems – major CRMs, meeting recorders and revenue-intelligence tools – on a schedule, into cloud storage. Stateless and independently scalable, so a slow sync never touches the app.
- OAuth
- 7 sources
- GCS
Validate, Load & Clean
A three-stage pipeline validates every incoming file, quarantines the failures, loads the good data to a raw store, then cleans it into a normalized shape – so downstream AI never sees malformed input.
- Event-driven
- PostgreSQL
Summarize & Vectorize
Transcripts are summarized and embedded into a vector store, isolated per organisation. This is the substrate that makes semantic search and retrieval possible.
- Embeddings
- pgvector
- Multi-tenant
Insight Engine
A batch pipeline runs retrieval-augmented generation and clustering to turn the vectorized data into the insights, themes and signals that appear on the dashboard, persisted to an analytics database.
- RAG
- Clustering
Agent Layer
Four purpose-built AI agents – insight Q&A, content generation, data analysis and executive briefings – stream answers to the app, and an MCP server exposes the same tools to external AI clients. Every query scoped to one organisation.
- Agno
- MCP
- SSE streaming
Multi-Tenant App
The web application: organisation-level data isolation, a five-role permissions model, subscription billing, and a modern React front end – the surface everything else feeds.
- FastAPI
- Next.js
- Stripe
- Next.js 15 & React 19
- FastAPI & SQLAlchemy 2.0
- PostgreSQL & pgvector
- Agno agents & MCP
- Terraform
- Hetzner & GCP
- Stripe billing
More Than Hands on a Keyboard
Embedded tech lead means three jobs at once: architect, builder, and the technical gate before production.
Architect
01
System design from first principles: the service boundaries, the data contracts between them, the choice to split ingestion, batch AI and interactive AI into separate scalable pieces. Decisions that still hold a year of growth later.
Builder
02
Hands on the keyboard across the whole stack – AI pipeline, backend, infrastructure, front end – and the reviewer who sets the quality bar for everyone else's. Plus the un-glamorous wins, like consolidating three drifting data repos into one clean ingestion monorepo.
Release Authority
03
Formal technical sign-off: nothing reaches production without it. Dependency and risk analysis across a release, migration and rollout planning, and the final go / no-go call – the safety valve between "done" and "deployed".
Built Like It's Meant to Last
A one-person prototype and a platform a team can safely ship to every week are different things. This is the second one.
Infrastructure as Code, Multi-Cloud
The whole estate is defined in Terraform across production and a parity sandbox, spanning two cloud providers, with disaster recovery in its own region. Environments are reproducible, not hand-tended – and a new one is a config change, not a weekend.
- Production + sandbox parity from the same modules
- Cross-region disaster recovery
Tested End to End
A dedicated Playwright suite exercises the real user journeys against both sandbox and production, so regressions are caught before a release reaches the sign-off gate.
CI/CD & Shared Workflows
Automated build, test and deploy across every repo, with a shared library of reusable pipeline steps so seventeen services stay consistent instead of each reinventing its own.
Multi-Tenant by Construction
Organisation isolation runs through every layer – storage, embeddings, agents, app – with role-based access control, so one customer's data can never surface in another's.
A Real Process
Definition of Done, a QA stage, and release management with formal technical sign-off – lightweight governance that lets a team move fast without breaking production.
Need Technical Leadership for Your AI Product?
We embed as your tech lead: architecture, hands-on engineering across the whole stack, and the judgement to say when a release is actually ready. From first commit to production, and the year after.
Open for new projects
See starting prices for this kind of work, or tell us what you're building and we'll reply within 24 hours. Rather use one of our products, like Tvrtko.ai, Moj Kolega or Titlomat? Those are open too.