DRAFT
The PodcastIndex feed record exposes 23 episodes for The Fund AI Pod, a weekly show about AI use in the funds industry. The earlier queue named EP12, EP14, and EP19 but did not promote the feed’s later episodes or its platform-hosted SRT transcripts. This pass recovered all 20 Spotify-hosted transcript files for EP03–EP22, then downloaded and transcribed EP01, EP02, and EP23 locally because no platform transcript URL was returned for those episodes.
The SRTs are platform transcripts used for timestamp navigation. They are not manually verified verbatim transcripts. Proper names, product spellings, and quantitative claims should be checked against the audio before quotation.
Episode URL: The canonical episode URLs are listed in the recovered episode inventory below.
Source files: sources/13-multimodal-sources/fund-ai-pod-feed-expansion-2026-raw.md; research/13-multimodal-sources/fund-ai-pod/raw/; PodcastIndex feed record and episode metadata; the EP23 MP3 enclosure.
Credibility: HIGH for feed identity, episode identity, dates, and returned transcript URLs; MEDIUM for guest-reported product, architecture, employment, and AUM claims; LOW for any claim that would require an unprovided benchmark, contract, or independent outcome audit.
Verification: Feed and episode metadata were resolved through PodcastIndex on 2026-08-18. Twenty Spotify SRT files were fetched from the returned transcript URLs. EP01, EP02, and EP23 were downloaded from their PodcastIndex enclosures and transcribed locally with MLX Whisper large-v3-turbo. All 23 transcript artifacts were checked for expected duration and substantive content. Claims are timestamped to the SRT and kept separate from independent verification.
Evidence Boundaries: These episodes do not establish complete model inventories, training data, customer permissions, live trading authority, out-of-sample performance, or AI-attributed returns. Vendor claims, named-practitioner accounts, and historical personnel statements remain distinct evidence classes.
Recovered episode inventory
| Episode | Date | Guest and affiliation on the publisher record | Local artifact |
|---|---|---|---|
| EP12 — William Wu, Menos AI | 2026-03-05 | Junchen (William) Wu, CEO, Menos AI | raw/ep12-william-wu-menos-ai.srt |
| EP14 — Alex Benke, Ridgeline | 2026-03-18 | Alex Benke, Head of Artificial Intelligence, Ridgeline | raw/ep14-alex-benke-ridgeline.srt |
| EP20 — Pat Starling, FactSet | 2026-05-20 | Pat Starling, AI Foundry leadership, FactSet | raw/ep20-pat-starling-factset.srt |
| EP21 — Shu Bai | 2026-05-27 | Hedge-fund portfolio manager; the episode describes prior/current exposure to Barron Capital, Davidson Kempner, and Balyasny | raw/ep21-shu-bai.srt |
| EP22 — Declan Sheehy | 2026-06-10 | Former CTO and COO, HSBC Alternative Investments; consultant and author at recording time | raw/ep22-declan-sheehy.srt |
| EP23 — Brad Olesen, Benzinga | 2026-06-23 | Chief Content Officer, Benzinga | raw/ep23-brad-olesen-benzinga.mp3; raw/ep23-brad-olesen-benzinga.srt |
Feed-wide transcript recovery
The feed is now fully represented at the transcript-navigation layer. EP03–EP22 have platform-hosted SRTs; EP01, EP02, and EP23 have local MLX Whisper SRTs generated from the publisher audio enclosure. This closes the feed-level capture gap but does not make every episode a firm-specific AI disclosure: several are adjacent vendor, governance, legal, or enterprise-transformation context.
| Episodes | Capture status | Local artifacts | Interpretation boundary |
|---|---|---|---|
| EP01–EP02 | Local ASR recovered from PodcastIndex enclosures | raw/ep01-diana-dinis-bunch.srt; raw/ep02-bernard-hanratty-funds-industry-ai-adoption.srt |
Product, governance, and board-level practitioner accounts; proper nouns and quantitative claims require audio or first-party confirmation. |
| EP03–EP11 | Spotify SRTs recovered from PodcastIndex transcript URLs | raw/ep03-...srt through raw/ep11-...srt |
Platform transcripts suitable for navigation and paraphrase, not unverified verbatim quotation. |
| EP12–EP22 | Existing and newly recovered Spotify SRTs | Existing canonical short names plus raw/ep13-...srt, raw/ep15-...srt through raw/ep19-...srt |
The five previously archived episodes remain canonical; duplicate downloads were removed. |
| EP23 | Local ASR recovered from PodcastIndex enclosure | raw/ep23-brad-olesen-benzinga.srt and .mp3.meta.md |
Local ASR has proper-name and noise limitations; use the audio for quotation. |
Additional timestamped evidence from the recovered feed
| Episode and time | Observable signal | Evidence boundary |
|---|---|---|
| EP01, 10:47–14:06 | Diana Dinis describes a three-layer sequence for private-markets automation: clean data, governance/policies, then AI. She identifies document extraction and categorisation as the current practical use case, with a human comparison step between source documents and extracted fields. | Named VP Product account at Bunch; no independent accuracy audit, customer list, or production-control documentation. |
| EP01, 14:13–15:33 | Bunch describes V7 for document extraction, standardised prompts, LPA-specific workflows, and extraction of fees, hurdles, carry, KPIs, and other terms into a database. | Vendor/tool disclosure by a practitioner; the transcript does not establish contract scope or whether every workflow is in production. |
| EP02, 00:00–00:08 | Bernard Hanratty rejects replacing a company secretary with an autonomous agent and says the role remains. | Short opening statement from a local ASR transcript; treat as a governance posture, not a universal rule. |
| EP02, 04:23–07:35 | Hanratty describes board exposure to AI, the EU AI Act, and his view that funds-industry use cases generally require careful risk classification rather than treating regulation as a blanket reason to stop. | Personal board and industry-advocacy account; not legal advice or a regulatory determination. |
| EP03, 00:00–00:20 and 53:40–55:50 | Jeb Altonaga is introduced with prior Citadel and Northern Trust experience; the episode later describes an RFP agent that deconstructs questions and drafts responses from internal material. | Historical personnel clue and practitioner/vendor workflow discussion; no current Citadel deployment or performance evidence. |
| EP06, 34:00–36:40 and 57:00–58:40 | Duncan Cooper discusses data governance, data products, and agentic workflows including knowledge-management and RFP agents. | Former CDO / fund-services practitioner account; no named client deployment or measured outcome. |
| EP09, 00:00–00:19 | Toby Glaysher states that FINBOURNE’s prospectus agent can turn a 150-page prospectus into a 1,000-field fund record in three minutes, with a claimed 95–98% accuracy. | First-party vendor claim in a podcast. The figure is not independently benchmarked here; the transcript says confidence flags and human review remain part of the flow. |
| EP09, 12:31–15:52 | FINBOURNE describes an LLM routed through MCP servers to permission-traceable, auditable APIs, with confidence flags for human review during fund setup. | Vendor architecture description; no security review, customer configuration, or independent control test. |
| EP09, 20:21–21:59 | FINBOURNE describes rebalancing and reconciliation agents. One production example is described as covering about 2,500 accounts across roughly 25 custodians in under an hour with one employee. | First-party customer example; account count, timing, and staffing are not independently audited in this pass. |
| EP10, 00:00–00:18 | Hojun Choi says an AI agent can run five to ten investment hypotheses in 30 minutes to an hour, compared with hours of manual work. | LinqAlpha product/practitioner claim; no task definition, benchmark protocol, or investment result is supplied. |
| EP10, 01:55–05:54 | LinqAlpha describes a central intelligence layer over broker emails, newsletters, decades of internal notes, research, and market data; its system uses specialist agents, a finance ontology, and a proprietary multi-agent architecture. | Vendor architecture account; no customer data permissions, model inventory, or independent evaluation. |
| EP10, 09:21–09:51 and 12:20–15:18 | The episode describes multi-agent red-teaming, normalized data, roughly 30 agents assembled as reusable blocks, MCP-native integration, coding agents for structured data, and portfolio-construction or quantitative-analytics workflows. | Speaker-reported product design; “30 agents” and the described workflow are not independent adoption or performance evidence. |
| EP17, 18:39–19:31 and 37:50–38:03 | Alex Dunegan describes Gemini for data-science prototyping, Claude Code for daily workflow, Cursor for model switching, and a GitHub workflow with rigorous code review. He says a human approves the final step before execution. | Small currency-management firm account; tool use and control posture are public claims, not a complete production architecture or return attribution. |
| EP18, 00:00–03:15 and 06:20–13:20 | Simmons & Simmons describes its proprietary Percy AI tool, private-cloud operation, legal research and document workflows, model testing, and human review boundaries. | Professional-services workflow context; not evidence of hedge-fund deployment or an investment model. |
Newly discovered adjacent podcast source
The title-blind PodcastIndex search also surfaced Hedge Fund Huddle — “Trusted news in the age of AI: How hedge funds find signal in the noise”, published 2026-08-04. It was not part of the original feed queue. The episode names Vik Bansal, Systematic Portfolio Manager at Centiva Capital, alongside LSEG speakers. It has a local MP3 and SRT capture at research/13-multimodal-sources/hedge-fund-huddle/raw/; the detailed source ledger is hedge-fund-huddle-trusted-news-ai-2026-raw.md.
| Time | Observable signal | Evidence boundary |
|---|---|---|
| 03:37–07:26 | Bansal describes trials of many datasets and news sources, a preference for differentiated properties such as local-language modeling or speed, vendor simulation evidence, and review by two experienced data engineers/developers. He describes human-led evaluation because vendors can handle corporate actions and historical mappings differently. | Named systematic-portfolio-manager account at Centiva; no dataset names, contracts, signal library, or independent process audit. |
| 07:56–09:53 | LSEG says hedge funds and asset managers are using AI for cross-checking and cross-referencing news and sentiment, with ongoing oversight rather than one-time data approval. | Vendor client-facing account; no client identities, adoption denominator, or measured error reduction. |
| 17:54–21:37 | The panel names satellite, shipping, sentiment, video, image, and transcript data as alternative-data modalities, then connects new-source adoption to ingestion, cloud, AI-model, MCP, and token-consumption costs. | General client/vendor discussion; no named hedge-fund strategy, cost model, or return attribution. |
| 24:15–26:12 | Bansal says Centiva uses LLM-based sentiment models, including local-language cases, but does not use AI to create signals on its own. An AI-generated signal would be treated like a data provider and would need out-of-sample checks. He also describes AI-assisted code prototyping and a roughly 40% speed-up in one live-trading process after AI identified an unknown bottleneck. | Direct practitioner account, but signal methodology, sample, baseline, timing, and speed measurement are not disclosed. |
| 28:44–30:42 | LSEG describes AI-assisted checking over very large content and options datasets while retaining human validation because an apparent spike may be a real block trade or auction event. | First-party workflow description; no error-rate or control-test evidence. |
Newly discovered buyside podcast/video source
The wider web search surfaced Fundamental Edge’s “Stoic Point’s Raj Shah: AI and the Lean Hedge Fund”, with a public YouTube recording and public captions. Raj Shah is described as co-founder of Stoic Point and a former Light Street partner. This source was not in the earlier queue. The canonical caption artifact is research/13-multimodal-sources/invest-with-ai/raw/stoic-point-raj-shah-lean-hedge-fund.en.srt.
| Time | Observable signal | Evidence boundary |
|---|---|---|
| 12:00–15:00 | Shah describes separating deterministic screening from non-deterministic qualitative work: starting with a quantitative Bloomberg/EQS universe, then handing a bounded list to tools such as Portrait or AlphaSense for qualitative screening. He describes internal work to combine those strengths. | Named fund co-founder account; no system diagram, vendor contract, or independent completeness test. |
| 15:39–21:08 | The episode describes a fragmented workflow across PDFs, spreadsheets, websites, Bloomberg, AlphaSense, MCP-connected tools, and agents. A “meta-screen” aggregates roughly 50 internal screens and flags companies appearing on multiple screens. A portfolio file can feed an agent that asks Portrait for peer companies, then creates an AlphaSense monitoring list covering news, expert networks, and SEC filings. | Public interview account; the described workflow is not evidence of a realized return or complete production implementation. |
| 21:37–22:10 | Shah gives a monitoring example in which AlphaSense surfaced a luxury company’s commentary about Lux Experience after Portrait identified it as a peer, leading to an unexpected monitoring alert. | Anecdotal workflow example; no counterfactual, trade record, or attributable P&L is supplied. |
| 25:10–27:27 | The speakers describe turning a PM’s screening, research, monitoring, scorecards, and investment letters into an intern guide or “sparring partner,” including a low-stakes agent review before a human pitch. | Described training and review pattern; no firm-wide policy, evaluation set, or personnel impact measure. |
| 27:31–29:20 | Shah calls Excel-plus-AI a significant surprise in hedge-fund work and describes Claude/Codex integrations for model-building and analysis. | Practitioner tool-use account; exact products, permissions, data handling, and code-review controls are not fully specified. |
| 35:50–36:45 and 47:27–48:40 | The episode frames AI ROI in P&L terms but says attribution is difficult. Shah then describes a possible division in which human judgment remains central to idea generation while AI supports portfolio management, factor-risk adjustment, and related controls. | Opinion and operating hypothesis, not measured attribution or a disclosed Stoic Point policy. |
Timestamped evidence map
| Episode and time | Observable signal | Evidence boundary |
|---|---|---|
| EP12, 12:42–14:40 | William Wu describes three agent surfaces: investment research and idea generation, position sizing through a product he calls “Voice Scoring,” and portfolio/risk management. | Menos AI’s account of its own product; no model card, training corpus, validation sample, or investment attribution. |
| EP12, 14:40–16:45 | A unified research hub is described for analyst email, third-party research, bank research, estimates, filings, global macro, and credit workflows. | Guest account; no named customer, licensing terms, coverage denominator, or accuracy audit. |
| EP12, 17:15–20:34 | The guest distinguishes a general LLM from a domain workflow that structures unstructured research, uses an agent to write code against quantitative data, and retains source/collaboration context. He describes this as deterministic and explainable. | Product claim, not an independent test. The raw transcript contains ASR errors; avoid quoting “no hallucination” language without audio review. |
| EP12, 22:43–24:15 | “Voice Scoring” is described as examining the logic, reasoning, conviction, and behavior patterns in research or trade ideas to help separate skill from luck. | The discussion does not disclose the label construction, time horizon, feature set, model family, or out-of-sample result. |
| EP12, 24:40–26:20 | Operational examples include repetitive data cleaning, statistical tests, and trade reconciliation, with the stated objective of returning analyst time to analysis and decision-making. | Vendor/practitioner account; no measured time saving or production-control evidence. |
| EP14, 08:25–10:20 | Alex Benke describes a Ridgeline assistant grounded in help documentation, followed by agents understood as digital coworkers with observation loops, approval pauses, and a trade-compliance workflow that gathers email, historical violations, audit trails, and rule changes. | Platform-practitioner account; no customer deployment list, reliability result, or independent control test. |
| EP14, 13:24–17:39 | Agents are described as inheriting user-like permissions and portfolio access controls. Dashboards, notifications, and audit trails expose workstream status and human action points. | Description of platform design; no configuration, penetration test, or customer-specific permission audit. |
| EP14, 25:18–28:12 | Ridgeline describes internal model gateways, a developing evaluation framework for comparing models by user experience and cost, dynamic routing as an R&D topic, and experiments with open-source models. | No model list, scores, routing policy, or deployment result is disclosed. The transcript contains a likely model-name ASR error. |
| EP20, 01:20–06:23 | Pat Starling describes FactSet’s AI product/data strategy and an AI Foundry intended to coordinate work at the data, workflow, and AI-research layers. The episode names Kate Stepp as Chief AI Officer and describes a top-to-bottom stack review. | Executive interview and platform transcript; no organizational chart, budget, or independent delivery measure. |
| EP20, 07:41–08:42 | FactSet describes MCP access to FactSet data through Claude, ChatGPT, and Gemini, plus an ecosystem of roughly 20 AI partners. Publicly named examples in the episode include Portrait Analytics and Finster AI. | Speaker account and partnership references; scope, data entitlements, customer usage, and commercial terms are not disclosed. |
| EP20, 09:35–12:05 | The episode describes transcript assistance, the Mercury universal chat surface, internal private LLMs, pilots with Microsoft/Copilot and other providers, and experimentation with Claude Code. | Product and internal-use account; no adoption denominator, model inventory, or evaluation result. |
| EP20, 14:55–16:20 | Research workflows are described as combining news, broker research, expert calls, management meetings, FactSet data, and external tools through MCP. | Workflow illustration; it does not establish use by a named hedge fund or a trading outcome. |
| EP20, 25:18–31:52 | FactSet describes model gateways, cross-model evaluation, cost-aware model selection, open-source model experiments, data lineage, provenance, QA, source links, and grounding answers in supplied facts. | First-party design and operating description; no reproducible benchmark or customer audit is supplied. |
| EP21, 00:59–02:45 | Shu Bai describes a public-markets path through Barron Capital, Davidson Kempner, and Balyasny, including Asia technology investing and a broader global technology mandate at the time described. | Episode self-report, date-scoped to the recording; it does not establish a firm-wide Balyasny AI program. |
| EP21, 03:15–06:20 | Bai frames AI investing around repricing, applications, sustainable cash flow, and the difficulty of assigning a three-to-five-year terminal value when model capabilities and company economics move quickly. | Investment thesis discussion, not evidence of a deployed model or proprietary signal. |
| EP22, 01:06–01:14 | Declan Sheehy says he spent 18 years with HSBC Alternative Investments and describes growth from $200 million to $31.6 billion across hedge funds and private equity. | Personal account in a podcast transcript; AUM history needs independent filing or company-source verification before use as a league-table fact. |
| EP22, 07:17–11:59 | Sheehy argues that data pipelines, a consistent data model, structured/unstructured integration, and a cross-functional data view precede useful GenAI; he sketches a central intelligence layer over research, portfolio management, operations, compliance, risk, and client service. | Practitioner architecture opinion; no HSBC implementation artifact or measured outcome. |
| EP23, 00:42–01:24 | Brad Olesen describes a career spanning trading, Bloomberg, Seeking Alpha, an RIA startup, and Benzinga. The publisher metadata, rather than the automatic transcript, is used for the proper name and current title. | Date-scoped personnel evidence; the recording does not expose Benzinga’s complete AI or data organization. |
| EP23, 02:20–03:09 | Olesen describes Benzinga as a news and data provider distributed through brokerage platforms including Robinhood, Fidelity, and Schwab, with earnings results, calendar events, and other data also supplied to prediction-market platforms. | Interview account; individual distribution agreements, schemas, data rights, and customer usage are not disclosed. |
| EP23, 05:00–07:02 | Benzinga Pro is described as combining fast-moving global information with U.S. stock coverage produced by Benzinga staff, Wall Street contacts, analysts, and in-house analysis. The interview frames speed as necessary but says context from a reporter remains useful after machines react to a release. | Practitioner description of a media workflow, not evidence of a specific hedge-fund feed, signal, or latency advantage. The local ASR contains a long repeated/noisy interval around 03:30–05:00 and should not be quoted. |
| EP23, 09:02–12:07 | Benzinga describes two AI lanes: customer-facing AI tools built through partnerships, and internal use for research and reporter productivity. Internal use is described as querying the existing content database for context and scaffolding, assembling data and productivity-suite information, and putting analyst-rating calendars and earnings data in front of reporters so they spend less time on menial research. The speaker also says Benzinga retained editorial investment and pulled back from formulaic automated content during the “AI slop” period. | Named executive/practitioner account. No partner list, model inventory, retrieval architecture, training data, adoption denominator, or measured productivity result is disclosed. |
| EP23, 12:47–14:01 | Olesen confirms a Perplexity partnership in which Benzinga supplies information that is made available through Perplexity’s AI platforms; he says he cannot discuss deal specifics. | Partnership existence is speaker-reported and licensing details remain undisclosed. This does not establish a custom model, exclusivity, or hedge-fund customer deployment. |
| EP23, 14:30–16:46 | Benzinga says users can access AI capacity around historical stock performance and Benzinga’s own coverage. The interview describes a highly skewed usage pattern with heavy users and dabblers, but gives no adoption percentage or usage denominator. | Qualitative adoption account; no product telemetry or evaluation result is supplied. |
| EP23, 18:00–20:17 | Benzinga says it launched an MCP in the prior year, has a longstanding licensing/API business, and is moving MCP and self-serve access toward B2C users. The guest describes institutional demand for data outside the portal and consumers building their own tools; a possible 4–5% conversion uplift is stated as an expectation, not a measured result. | Product and forward-looking commercial account. No MCP schema, authentication model, rate limits, permissions, customer count, or realized conversion uplift is disclosed. |
| EP23, 21:31–28:21 | Benzinga describes prediction-market partnerships and relationships in which it helps resolve markets, including questions of what actually happened. It says prediction markets are also used internally to gauge interest in financial-media topics, and describes a dedicated prediction-market coverage vertical. Benzinga is described as a data/content provider rather than an execution venue in the current setup. | The episode does not identify all partners, provide a resolution dataset, or establish trading, market-making, or investment use by Benzinga or any hedge fund. |
| EP23, 28:56–31:10 | For AI-assisted financial content, the speaker describes repeated checks, human editorial review, restricting AI to work that ideally is not market-moving or central to a trade idea, chunk-level double/triple verification, and a final human “smell test.” | Stated control design, not an independent audit. No error rate, incident log, model-risk policy, or review-coverage denominator is provided. |
| EP23, 37:12–40:04 | Benzinga frames generative-engine citations as downstream of traditional quality, authority, expertise, and comprehensiveness rather than a separate tactic. The guest says citation-driven referral traffic is currently small and expresses concern about AI summaries that reuse publisher work. | Executive opinion and traffic observation without a disclosed measurement method; not a general ranking of media strategies. |
What this closes, and what remains open
This pass closes a feed-registry gap and adds evidence lanes for investment research, position-sizing analysis, fund operations, system-of-record permissions, financial-data provenance, model evaluation, historical hedge-fund personnel lineage, financial-media AI, MCP distribution, and prediction-market resolution. It also surfaces a new partner-discovery lane: market-data and financial-media vendors disclose AI partners and data-delivery surfaces in practitioner podcasts even when the partner page is not indexed by hedge-fund terms.
EP01, EP02, and EP23 are verified in PodcastIndex metadata and now have locally archived MP3s and MLX Whisper SRTs. The local transcripts can misrecognize proper nouns and should be used for timestamp navigation and paraphrase, not verbatim quotation. EP03, EP06, EP09, EP10, EP17, and EP18 add specific workflow and control signals that were previously absent from the public evidence map. EP19, EP20, and EP21 also show why title-blind monitoring matters: the firm or role is often in the episode description rather than the title.
None of these sources establishes live trading authority, a complete model inventory, training-data rights, out-of-sample performance, or AI-attributed returns. The evidence remains separated into vendor claims, practitioner accounts, platform design descriptions, and historical personnel context.