Five tools collapsed into one agentic AI platform for PepTalk

Twistag rebuilt PepTalk's speaker-booking operations on four chained AI agents, collapsing five disconnected tools into one platform and indexing more than 3,000 speaker profiles as vector embeddings for semantic matching. PepTalk, a speaker booking marketplace, receives hundreds of unstructured speaker requests a month, and each one previously moved by hand between an inbox, a speaker database, a document editor, a CRM and a calendar. Twistag built an inbound parser that turns any request format into structured data, a semantic matching engine on Supabase with pgvector, a proposal generator on Amazon Bedrock with guardrails, and a follow-up agent that writes structured data back to the CRM.

Every PepTalk request arrives as unstructured text — a polished RFP, a two-line email, a phone call transcribed into notes — and each one requires extracting the parameters (event type, date, topic, budget, audience size), finding suitable speakers, assembling a proposal and managing the deal to close. That worked at lower volumes. At hundreds of requests a month it became a bottleneck, with the sales team spending more time on data extraction and cross-referencing than on the consultative work that closes bookings. Fragmentation made it worse: at least five systems touched each request, and a single booking required copying data between all five. Proposals sometimes went out with outdated speaker availability. For a speaker agency this is a revenue problem, not an operational one: a client who sends the same brief to three agencies typically books with whoever responds first with a relevant shortlist.

Twistag framed PepTalk's problem as an agent problem rather than an integration problem, on the reasoning that most of the manual work was structured reasoning rather than creative judgment, and that middleware connecting existing tools would only connect the existing friction.

The platform runs four chained agents. The inbound parser reads a request in any format and outputs a structured JSON object covering event type, date, topic, budget, audience profile, language requirements and constraints; it runs on Amazon Bedrock and flags ambiguity explicitly rather than guessing. The speaker matching engine is the analytical core: PepTalk's 3,000-plus speaker catalogue is stored as vector embeddings in Supabase using pgvector, capturing topic expertise, speaking style, audience fit, fee range, geographic availability, languages and past performance ratings, and the agent converts event parameters into a query vector and runs a multi-dimensional similarity search, returning a ranked shortlist with relevance scores and match reasoning. The proposal generator produces a customised document per request, with Bedrock prompt versioning for iteration and rollback and Bedrock guardrails checking every proposal for factual consistency, pricing accuracy and tone before it can be sent. The fourth agent manages the deal after the proposal goes out, triggering follow-up on engagement signals rather than rigid schedules and logging every interaction back into the CRM as structured data.

The review interface is where the agents get better. Twistag ran Argilla across all four agents as a continuous evaluation layer: every AI output is reviewable in a structured interface, and sales-team corrections feed directly into prompt optimisation. That makes improvement a normal part of using the system rather than a retraining project.

Stack: Amazon Bedrock, Supabase, pgvector, Argilla. Services: AI agents and multiagent systems, AI-native product engineering, agent operations and evaluation.

Five separate tools collapsed into one platform, with four agents running in production against a catalogue of more than 3,000 semantically indexed speaker profiles. The inbound parser removes the first 10 to 15 minutes of manual handling on every request and standardises the data before any downstream agent touches it. PepTalk's CRM now reflects reality, so the sales team makes decisions on accurate pipeline data. The team spends its time on relationship building and consultative selling rather than data entry, and handles higher request volume with faster turnaround. Twistag has not published a measured figure for the volume increase, the turnaround time, or booking conversion.

Work that looks like judgment but follows patterns is exactly the agent zone.