Junior knows which words matter

The sentence was perfectly respectable: “We experimented with malware.” In the engineering account published by Junior, the intended word was Milvus, a vector database. The transcription had produced something readable, grammatical and wrong. For a team investigating a business, that is an awkward combination. A garbled word invites a second listen. A plausible sentence can slip straight into the research.

The error hiding in plain English

Junior’s engineers benchmarked speech systems on real commercial due diligence calls, checking important entities and facts against human-verified transcripts. Leading systems scored 50-75% on that measure. The team used larger multimodal models and more conversational context, then added checks and fallback models to catch invented speech. Junior reports that entity accuracy reached 97% on its benchmark. That is a company test of a particular task, not a universal accuracy rating.

The distinction matters. A transcript can contain thousands of correct ordinary words while mangling the three businesses everyone came to discuss. The practical lesson travels beyond finance: measure the errors that change a customer’s decision.

Two friends, one familiar chore

Junior’s founders, CEO Dimitris Samouris and CTO Nico Moschopoulos, describe themselves on Wellfound as high-school friends from Greece. Samouris brought management-consulting experience in commercial diligence and growth strategy. Moschopoulos brought software experience at businesses including Airbnb and Petal. They had complementary views of a familiar chore: turning a large pile of conversations into something another person can use.

Today, Junior captures expert interviews and other deal conversations, structures transcripts and makes the material searchable. Its Alexandria library is designed to retrieve knowledge across a firm. The company reports use at 20 of the top 30 global private equity funds by assets under management and 18 of the top 20 strategy consultancies. Those are Junior’s adoption claims; they describe its target market as clearly as any slogan.

A spreadsheet with a memory

Consider the next job after an interview: comparing what different experts said about competing businesses. Junior’s Competitive Set organizes companies against tracked metrics, lets a researcher open the underlying questions and answers, and exports the comparison to Excel. Its help documentation also describes merging duplicate names and splitting entities that were wrongly combined.

These controls deserve attention. Experts do not always use consistent names. Neither does software. Giving a researcher a way to repair the record is part of making the analysis usable. The matrix fills from processed calls; a subject nobody discussed stays empty. A blank cell can be a useful instruction for the next interview.

The headphones problem

Some of Junior’s most instructive work begins in customer support. Engineer Noah Mack describes users putting on headphones while a web recorder depended on the laptop microphone hearing the other side through its speakers. The result: half a conversation. Others forgot to stop recording, leaving enormous files that took approximately 15-30 minutes of manual work to recover.

The team built a Chromium extension that captures tab audio and observes the tab’s lifecycle. Closing the recorded tab can start processing. This also offered an option for customers whose IT departments were reluctant to approve a native application. Junior treated repeated mistakes as product failures. That is a useful habit to copy: if several customers make the same mistake, inspect the interface before writing another reminder.

Let the human interrupt

Athena, Junior’s autonomous interviewer, takes an interview guide and conducts an expert call. When asked to add human takeover, engineer Shahrizod Bobojonov initially responded, “This is a stupid feature.” Why build an interviewer and then put the researcher back into the interview?

“This is a stupid feature.”Shahrizod Bobojonov, recalling his first reaction to human takeover

His account explains the change of mind. Discovering a weak interview afterward means discovering it when the expert’s time and the deal team’s time have already been spent. A supervisor can now watch, intervene and return control. Athena keeps listening during the interruption, so it can continue with the human exchange in context. Oversight becomes a way to try automation without surrendering the whole conversation.

The careers page describes weekly deployments and ownership from conception through delivery. Its team photographs include international offsites and, more memorably, laptops beside a swimming pool. The engineering posts supply the useful counterpart to those recruiting pictures: named people explaining particular decisions, including the ones they had to reconsider.

The last mile is a slide

Junior’s PowerPoint work exposes another mundane obstacle. Engineer Eoin Coulter describes asking a model to identify sections from screenshots, only to get hallucinated sections. Feeding it the slide’s individual shapes and metadata worked better. Screenshots remained useful alongside that structure for deciding where a quote should go.

The extension searches project calls and documents, then helps apply edits inside PowerPoint. That places Junior close to the output deal teams actually deliver. Its competitive argument is specialization: capture the conversation, retain its meaning and carry evidence toward the report. Buyers can judge that against the work required to assemble the same process from meeting software, spreadsheets and a general AI assistant.

What the purchase commits you to

Junior sells SaaS through order forms. Its fulfillment policy specifies a minimum one-year term for some subscriptions, while consumption-based services are invoiced monthly for prior usage. A sensible evaluation should therefore follow a real project: record representative calls, check difficult names, inspect the source behind comparisons and test the export someone will actually use.

The payoff depends on the material going in. Unrecorded conversations cannot populate the library. Thin interviews cannot support rich comparisons. Researchers still need to notice missing questions and challenge conclusions. Junior’s revealing choices acknowledge those conditions: repair the names, expose the evidence, and keep a route back to the microphone.