- Publisher: Wharton FinTech Podcast
- Episode: “Schonfeld CTO Tom DeBow: Technology and AI’s Impact on Modern Hedge Funds”
- Published: 2025-04-15; publisher metadata lists 42 minutes.
- Speaker named by the publisher: Tom DeBow, CTO, Schonfeld.
- Recovered audio: SoundCloud/RSS enclosure listed in the local source metadata.
- Local navigation artifact:
tom-debow-schonfeld-2025.srt, 589 ASR segments, approximately 42:05.
The local file is an MLX Whisper large-v3 transcription used for timestamp navigation. Proper names and product names require manual audio spot-checking before quotation. The notes below paraphrase the recording and retain the timestamp so the claim can be rechecked.
Source files: sources/13-multimodal-sources/audio/wharton-fintech/rss-75980369faf1-schonfeld-cto-tom-debow-technology-and-ais-impact-on-modern-hedge-funds.mp3.meta.md; research/13-multimodal-sources/wharton-fintech/raw/tom-debow-schonfeld-2025.srt.
Credibility: HIGH for dated publisher metadata, named CTO role, recovered audio, and timestamp navigation; MEDIUM for firm-reported adoption and workflow claims; LOW for any inference about returns, permissions, or current model ownership.
Verification: Publisher metadata and the local audio sidecar were checked on 2026-08-18. The SRT was generated from the recovered audio with MLX Whisper large-v3 and is retained for navigation, not as a verbatim quote source. Later Schonfeld first-party pages were used as a separate current-state cross-check.
Evidence Boundaries: This source does not establish a complete model inventory, training corpus, adoption denominator, evaluation results, live trading permissions, or AI-attributed returns.
Timestamped evidence map
| Time | Observable disclosure | Boundary |
|---|---|---|
| 05:00–05:38 | DeBow describes the “last mile” as gathering and normalizing data, coding, visualization, and the work that turns inputs into an investment idea or analytical conclusion. | A process description, not a measure of investment impact. |
| 10:36–11:16 | AI work is described as sitting within an Applied Technology Team working directly with users. | The interview does not provide the team’s headcount, reporting lines, or complete remit. |
| 11:52–12:44 | Curated data, data structuring, and entitlements are treated as part of the system; sensitive data is described as visible only to authorized users. | No access-control configuration, audit record, or data inventory is disclosed. |
| 18:17–19:47 | DeBow describes an internal Schonfeld GPT surface available through Slack, email, and direct APIs, with multiple model families, internal APIs, and vector databases connected to proprietary data. The examples include coding, email digests, portfolio analysis, and trade-break diagnosis. | The recording does not enumerate models, vendors, embeddings, retrieval permissions, or production approval gates. |
| 21:58–24:05 | Model choice is described as task-dependent. Examples include coding, first-pass data science, Python work, filings, research, and news synthesis. | No benchmark, task-level accuracy, latency, or cost result is supplied. |
| 24:32–25:00 | DeBow repeats a firm-reported regular-use figure of roughly two-thirds of the user population and says the firm attempted cost and time calculations. | No denominator definition, baseline, absolute spend, calculation, or outcome is disclosed. |
| 39:04–39:35 | The discussion turns to customized bots around user workflows, including shifting routine trade-break processing toward exception handling. | This does not establish unattended remediation or trading authority. |
Later first-party cross-check
The interview is an earlier, dated practitioner account. Later Schonfeld pages add separate public evidence:
- Inside Schonfeld’s FE AI Lab describes SchonAI, pilots, evaluation, and investment-research workflow surfaces.
- Schonfeld careers and opportunities lists public AI Training & Education Analyst, AI Strategy Analyst, AI Data Engineer, and related technology roles.
- Hannah Jiang Q&A provides a separate personnel and engineering-workflow signal.
These later pages are not backdated into the 2025 interview. Together, the sources expose dated architecture and operating-model signals, but not a complete current model inventory.
Evidence boundary
This source supports a dated account of internal AI interfaces, data preparation, entitlements, workflow integration, and human-plus-machine operating context. It does not establish model training data, model ownership, live trading permissions, evaluation results, adoption telemetry, or AI-attributed returns. The transcript is an automatic ASR artifact; use the audio for any verbatim quote or proper-name claim.