MVP Development in Berlin, Germany

HR Technology / AI

Two TU Berlin computer science PhDs had developed a language model fine-tuned on 50,000 anonymized German job interviews that could assess candidate responses for competency markers — predicting interview success with 78% accuracy (compared to 52% for unstructured interviews alone, per published meta-analyses). They had an EXIST grant (€150,000), a working model in a Python notebook, and a 12-month deadline to demonstrate a functional product with pilot customers. Their challenge: the model worked in a research context (feed it a transcript, get a competency score), but turning it into a product required: a web application where HR managers upload interview recordings or transcripts, an audio-to-text pipeline (most interviews are recorded, not transcribed), a results dashboard showing competency assessments, benchmarks, and hiring recommendations, GDPR compliance (processing interview recordings involves biometric data and requires explicit consent, data protection impact assessment, and legitimate interest justification), and integration with applicant tracking systems (Personio — the dominant German HR platform, and Greenhouse) to fit into existing HR workflows. They needed the product built in 10 weeks to meet their EXIST milestone and begin pilot conversations.

We built the assessment platform over 10 weeks: (1) Audio processing pipeline: HR managers upload interview recordings (audio or video). The system extracts audio, transcribes using Whisper (self-hosted to keep data in EU — no sending interview recordings to US-based APIs), and produces a structured transcript with speaker diarization (identifying interviewer vs. candidate). German language transcription required fine-tuning Whisper on German conversational speech — standard Whisper had 15% word error rate on German interview audio due to regional accents, technical terminology, and the code-switching between German and English common in Berlin tech interviews. Our fine-tuned model achieved 6% WER. (2) Competency assessment engine: the founders' model was wrapped in a FastAPI service. The transcript is segmented into candidate responses, each response is assessed against the competency framework (communication, problem-solving, technical depth, cultural fit — configurable per role), and results are aggregated into an overall assessment. The key product decision: we present assessments as evidence-based insights, not scores. Instead of "Candidate scored 7.2/10," the dashboard shows "Strong evidence of structured problem-solving: candidate described a systematic debugging approach in responses 3 and 7" — linking each assessment to the specific transcript passage that supports it. This was critical for GDPR compliance (the assessment must be explainable) and for HR manager adoption (HR professionals distrust black-box scores but trust highlighted evidence they can verify). (3) Results dashboard: per-candidate view showing competency assessments with transcript evidence, comparison view overlaying multiple candidates against the same competency framework, benchmark data ("this candidate's problem-solving evidence is stronger than 72% of candidates assessed for similar roles"), and exportable PDF reports for hiring committee review. (4) GDPR compliance layer: we conducted a Data Protection Impact Assessment (DPIA) before development (required when processing biometric data at scale). The system implements: explicit candidate consent workflow (the candidate receives an email explaining that their interview will be AI-assessed, with opt-out capability), data retention policy (recordings deleted after 90 days, anonymized assessments retained for benchmarking), right to explanation (candidates can request the basis for their assessment — the system generates an explanation document linking each competency rating to transcript evidence), and data residency (all processing on Hetzner Cloud in Nuremberg — no data leaves Germany). (5) Personio integration: HR managers using Personio can trigger assessments directly from the applicant's Personio profile. Assessment results appear in the Personio candidate record. The integration uses Personio's REST API with OAuth2 authentication.

EXIST milestone met: the product was demonstrated at the 12-month EXIST review with 3 active pilot customers — exceeding the milestone requirement of "functional prototype." The EXIST evaluators specifically noted the GDPR compliance as "exemplary for an AI product handling sensitive personal data." Pilot results: 3 Berlin companies (a fintech, a logistics company, and a consulting firm) piloted the platform over 8 weeks, assessing 280 candidates. HR managers reported: 40% time saving on interview evaluation (the AI pre-assessment replaced the 30-minute post-interview writeup that recruiters do after each interview), reduced disagreement in hiring committees (the evidence-based assessment provided a shared, objective starting point for discussion — reducing the "I liked/didn't like them" subjectivity), and 23% improvement in new hire performance ratings at 6 months (candidates selected with AI-assisted assessment scored higher on 6-month performance reviews than the historical average — early signal, not yet statistically significant at n=280). Seed funding: the EXIST grant was supplemented by a €750,000 seed round from a Berlin VC, closed 3 months after the MVP launch — the working product with pilot data was cited as the primary investment thesis. Development cost: €85,000 for the MVP. The startup is now hiring its own engineering team and expanding from the codebase we built — no rewrite needed because the architecture was designed for growth.