P_Rad

Ex MD and Head of High Yield Trading. Now, building AI workflows in finance.

[P_Rad @prashantrad]

The second use case is turning unstructured knowledge into structured analytical inputs.

The most valuable information you have in credit is not in any database. It is in your head, your call notes, your emails, and your years of conversations. The dealer who traded a bond knows how it actually cleared, who was reluctant, and what the real sentiment was behind the public narrative. The PM who has spoken with a CFO three times has pattern-matched something no filing captures. The analyst who has covered a sector for a decade has a mental model no terminal can replicate.

The problem is that most of it exists in unstructured form. And credit investors have a layer of access that equity investors do not. Dealer runs, broker commentary, secondary market flow. Bank meetings, lender presentations, restructuring advisor decks, amendment negotiations, rating agency conversations. None of it is public.

Access within credit is not equal either. Large holders get more management time, more dealer color, more involvement. But the real value is not in who gets the most access. It is in who builds and compounds on what they have. Most of what gets accumulated disappears. Bloomberg chats you never saved. Sell-side commentary you half-remember. IB pitches you read once. Syndication updates you skimmed during execution. Conversations with management that lived and died in your notebook.

AI makes that information more usable. It turns unstructured knowledge into structured inputs that feed the underwriting process: signals, flags, pattern matches, and historical context surfaced at the moment you need it.

When that knowledge feeds a credit decision, the output is different. Risks you would have missed. Patterns that connect across cycles. You are not just underwriting faster. You are underwriting better.

Structured knowledge compounds. Each deal adds to the base. Each cycle adds pattern recognition that feeds the next one. And unlike the expertise that retires with a senior analyst, it stays in the fund. Most credit professionals have spent careers accumulating this kind of knowledge.

Most do not yet know how much AI can do with it. That is the gap EigenStrategy is designed to bridge.

[P_Rad @prashantrad]

Here are two broad AI use cases in credit. This is the first.

Credit underwriting is not one workflow. A term loan, a bond, an ABL revolver, an ABS deal, and an infrastructure facility all require different analytical frameworks. If the routing is wrong, the analysis is wrong.

I built a credit underwriting orchestrator that reads a dataroom and routes the deal to the right framework automatically. It then layers in sector-specific analysis. An auto credit gets different questions than a retail credit. A gaming credit gets different questions than a healthcare credit.

That is what real underwriting looks like: a structure specialist alongside a sector specialist.

The workflows can also be tuned to how a specific practitioner thinks. A distressed investor approaches the same deal differently than a direct lender. A direct lender approaches it differently than an investment grade buyer. Same instrument, entirely different questions.

The hardest part, and the most valuable, is conflict detection. Not extracting what documents say. Finding where the real risk only appears when multiple facts are read together.

A lease abstract says a termination right is exercisable in November 2031 with 18 months notice. A lender underwriting a loan maturing in February 2031 needs to know the real decision deadline is May 2030, which falls inside the loan term. Reading the documents gives you two dates. Understanding the deal gives you the risk.

Any system can extract terms. The point is to underwrite. That is the judgment your best people carry in their heads, made available at scale.

The second use case is different. That is Post 2.

[P_Rad @prashantrad]

Finance is not short on interest in AI.

It is short on applications that fit actual investment workflows.

In credit, direct lending, structured credit, distressed, and the LP workflows around them, the bottleneck has never just been analysis. It has been structuring unstructured information. Credit agreements, deal documents, portfolio reporting, transcripts, internal research, and proprietary material across systems all had to be read, interpreted, and converted into something usable.

For years, a skilled professional was the structuring layer. The limiting factor was cost and capacity.

That constraint is starting to lift. The opportunity is not AI for finance in the abstract. It is AI applied to workflows where unstructured information has historically limited speed, consistency, and depth of judgment.

My background is credit, from investment grade through stressed and distressed. The last year has been spent building in that gap.

More here: https://t.co/fhg60S8R9t

[P_Rad @prashantrad]

I wanted local open models to work for credit analysis.

They are not there yet. I tested Gemma 3 4B locally on an M1 Pro against two credit datarooms: one structured loan analysis, one multi-document deal with lease abstracts, financials, borrowing base detail, and construction data.

On the structured task with an explicit checklist, it caught 15 of 18 known flags: 83%, roughly in line with the 12B model. Zero API cost. Data never left the device. That was promising.

Then the open-ended task: find conflicts across a heterogeneous document set. It scored 2%. It missed a termination right in an executed lease. It misread the advance rate. It hallucinated DSCR figures.

Fine-tuning on 37 examples made it worse. Truncated training examples taught format mimicry, not reasoning. Performance dropped 11 points.

The gap between those two results is the actual finding. Local models follow instructions. They do not yet reason reliably about what to look for.

Cloud LLMs still have a significant edge on real credit work. That matters because the institutions with the strongest demand for on-premises AI, banks, credit funds, and asset managers with real data obligations, are exactly the ones where that reasoning gap cannot be papered over.

The capability will get there. It is not there today.

[P_Rad @prashantrad]

I spent 17 years trading and managing books and businesses across credit.

I should not know what a context budget is.

But I do. And that tells you something about where we are with AI in financial services.

Token costs have fallen dramatically over the last two years. More providers, open-source competition, better inference efficiency, and tokens moving closer to commodity pricing.

That is real progress. It is one reason adoption is accelerating.
The problem is what happens next.

Costs fall, usage scales to meet them. Firms deploy frontier models for every task. Context windows balloon. Agents run when they should not. The per-unit cost drops, but the total spend does not follow.

This is the blowtorch problem.

Not that the gas is expensive.

That you are using the wrong tool.

I built a token optimizer to audit exactly this. Every skill, agent, and workflow mapped against what it actually costs, and whether it needs the model it is running on.

The discipline is not complicated: right model for the right task, hard context limits, regular audits of everything deployed.

Tokens becoming cheap does not mean you stop thinking about how you use them.

It means everyone else stops.

And that is the advantage.

[P_Rad @prashantrad]

The story being told about AI in finance is about what it takes away. I want to tell a different one.

The fear in credit, especially for analysts and associates, is displacement. That AI does their job and they become redundant.

I spent 18 years building judgment about credit risk. Which questions to ask first. Where deals break. What conflicts in the documents matter, and which ones are noise.

That judgment lived in my head and nowhere else. Over the past several months I encoded it. I built a set of skills that ask the same questions, in the same order, that I learned over nearly two decades. The output is structured, sourced, and calibrated the way I would have done it at the desk.

What that means for a 28-year-old analyst is not what most people assume.

It means they now have access to the same risk identification framework that took me until my late 30s to build. They can interrogate a credit agreement the way a senior PM would, not from a standing start.

That is not displacement. That is compression of the learning curve.

The analyst who embraces these tools moves up the curve faster. They start doing work that used to require 15 more years of pattern recognition.

[P_Rad @prashantrad]

After 17 years in credit trading, I left RBC knowing I wanted to build something different. I just did not know exactly what yet.

When I started working seriously with AI, the capability was obvious. So were the failure modes. The technology is powerful, but most of the mistakes show up at the intersection of model capability and domain judgment.

That is what led me to start EigenStrategy. The name comes from linear algebra. An eigenvector stays true to its direction even as the space around it changes. That felt like the right metaphor for what I wanted to build.

Investment firms are going to incorporate AI into their workflows. The question is whether they do it as a thin productivity layer or as a real extension of how analysts, PMs, traders, and operators make decisions.

What gets lost in most implementations is domain expertise. A technology team can deploy a model. They cannot tell you which covenant provisions matter in a stressed credit, how a trader actually uses positioning data, or why the first question in a liability management situation is not the spread but the indenture. That knowledge lives with practitioners, not engineers.

EigenStrategy exists to close that gap. Since the beginning of the year, I have been working with funds and financial services firms on exactly this. Now I am making it official.

[P_Rad @prashantrad]

Every AI disruption argument in finance starts at the analyst-to-associate layer. That gap is already closed.
It's table stakes now.

The more interesting question is whether AI meaningfully threatens the domain expert layer. My read: less than the current discourse implies, and for a structural reason.

The decisions that actually move risk in institutional finance don't get written down. How a specific management team behaves at covenant breach, which structuring angles a particular desk will accept, what a counterparty's real walk-away point is: none of that circulates through anything AI trains on. It lives in deal experience and institutional memory.

There's a growing body of ex-practitioner writing online, and some of it is genuinely good. But it captures the fraction that got written up. The most proprietary judgment is proprietary precisely because it never enters the system.

The exposed layer is not defined by title. It is defined by how much of the work can be documented, searched, reproduced, and checked.

The firms most exposed are the ones whose value-add is information processing, but who still describe it as judgment.

[P_Rad @prashantrad]

This week I built a credit market intelligence tool in about two hours.

It pulls HY and IG spreads from FRED, primary dealer inventory from the New York Fed, and live EDGAR filings for covenant amendment volume and going concern flags. Then it synthesizes all of it for about three cents a run.

Today's output: high yield index OAS sits at the 13th percentile of its 5-year range. But CCC spreads have widened 23 basis points over the past three months while BB and B have been flat to tighter.

In other words: the index headline is masking quality dispersion. The market is pricing the BB and B complex for near-perfection while quietly repricing the bottom of the stack in the opposite direction.

Six months ago, building something like this as a solo project would have been borderline ridiculous. The data plumbing alone would have required institutional-style infrastructure.

And the synthesis of technical market signals with filing-based stress indicators usually lived in someone's head, if it happened at all, across multiple desks and systems that did not talk to each other.

What struck me is that once you have real-time data access, this becomes a continuous loop: credit spreads, dealer positioning, filing-based early warning signals, and equity market context all running side by side.

That kind of integrated view, fundamentals and technicals, fixed income and equities, in one pass, is tooling I would not have had access to even when I was running a trading desk.

And I can do that right now.

[P_Rad @prashantrad]

This week I built a credit market intelligence tool in about two hours.

It pulls HY and IG spreads from FRED, primary dealer inventory from the New York Fed, and live EDGAR filings for covenant amendment volume and going concern flags. Then it synthesizes all of it for about three cents a run.

Today's output: high yield index OAS sits at the 13th percentile of its 5-year range. But CCC spreads have widened 23 basis points over the past three months while BB and B have been flat to tighter.

In other words: the index headline is masking quality dispersion. The market is pricing the BB and B complex for near-perfection while quietly repricing the bottom of the stack in the opposite direction.

Six months ago, building something like this as a solo project would have been borderline ridiculous. The data plumbing alone would have required institutional-style infrastructure.

And the synthesis of technical market signals with filing-based stress indicators usually lived in someone's head, if it happened at all, across multiple desks and systems that did not talk to each other.

What struck me is that once you have real-time data access, this becomes a continuous loop: credit spreads, dealer positioning, filing-based early warning signals, and equity market context all running side by side.

That kind of integrated view, fundamentals and technicals, fixed income and equities, in one pass, is tooling I would not have had access to even when I was running a trading desk.

And I can do that right now.

#CreditMarkets #HighYield #FixedIncome #AI #InvestmentManagement

[P_Rad @prashantrad]

Everyone building with AI started from zero.

The people who look far ahead just started earlier.

There's two kinds of AI content right now: people deep inside it, and everyone else watching from the edge.

That second group is much larger than it looks. And a lot of them feel like the gap is permanent. The models change too fast. The vocabulary assumes prior knowledge. The people posting about it seem like they've been doing this forever.

A year ago, I was in that second group. Curious, but unsure where to begin.

What I learned: the starting gap was the real gap. Not the knowledge gap.

AI is the first technology I've worked with where being wrong is part of the interface. A weak prompt gives you feedback. A bad output gives you something to steer against. Every attempt makes the next one less vague.

The pace of change feels like a reason to wait. It is actually the reason to start. Everyone is new to the current models.

The gap is smaller than the feed makes it look.

But only if you begin.

[P_Rad @prashantrad]

@LeveredThoughts agreed. its the first test.

[P_Rad @prashantrad]

When BDC prices diverge from NAV, LPs face a secondary market decision: sell at a discount, add, or hold.

The information needed to answer that question sits with the GP, not the LP. Loan-level performance, amendment history, PIK insertion rates, covenant headroom by position. What the LP typically has is a fund-level NAV mark, an audited annual, and a quarterly letter that cleared compliance before it arrived.

That gap was manageable when private credit was a hold-to-maturity asset class. Secondary activity at scale was not part of the design. The information architecture reflects that.

BDC repricing is stress-testing it.

LPs are being asked to make secondary market decisions using portfolio-level data designed for a different question: do I want exposure to this manager? Not: do I want this specific portfolio at today's secondary price?

ICE and Apollo announced infrastructure intended to improve that. Directionally right. But GP adoption, data standardization, and LP access all take time. That gap does not close at the speed of secondary market moves.

Until it does, secondary decisions run on price action, peer-fund mark comparisons, and manager reputation.

Enough to answer one question: do I want exposure to this manager?

It doesn't answer the question BDC repricing is forcing: do I want this specific portfolio at today's price?

And it says very little about exit: which positions have secondary depth, at what discount, and through which intermediaries.

Three questions. One dataset designed for the first one.

[P_Rad @prashantrad]

Elon Musk has argued publicly that human-level AI could arrive within a year. This week he testified in a trial that puts a related but different question on the table.

OpenAI is being sued over its conversion from a non-profit to a for-profit structure. The specific legal theory is narrow. It applies to OpenAI's structure, not to every AI lab.

But the case is doing something broader than settling a charter dispute. It is the first public test of whether a safety-first mission can hold once commercial pressure becomes large enough.

Other AI labs are structured differently. The underlying tension is not.

If Musk's timeline is even half right, that tension gets resolved fast, under pressure, in public, with a lot of money at stake.

Whether "responsible AI" is a structural commitment or a marketing claim is one of the more consequential governance questions of the next few years.

The OpenAI trial is the first data point.

[P_Rad @prashantrad]

#AI #CloudComputing #Enterprise #Investing

[P_Rad @prashantrad]

Google, Microsoft, Amazon, and Meta are spending on AI infrastructure at a pace that has no precedent in corporate history. Google's stock jumped 7% after Q1 earnings this week. The market read it as confirmation that the adoption flywheel is turning.

It is worth slowing down on what Google's quarter actually shows. Cloud grew 63% year-over-year because builders (companies consuming compute to train models and ship AI products) have insatiable demand. Google said it was capacity constrained: revenue would have been higher if supply could keep up. That is strong evidence of builder demand. It is not the same thing as enterprise deployment at scale.

Those are different things. Large organizations running AI across real workflows still face friction: regulatory exposure, cost uncertainty, internal governance, and plain inertia. According to reporting in The Information, Uber's CTO said their AI tools budget for 2026 was exhausted by April, at a company with $3.4 billion in R&D. Cost surprises at that scale slow adoption decisions.

There is a coherent scenario where the infrastructure gets built on schedule and the monetization of broad enterprise AI takes five years instead of two. In that scenario, the companies funding the buildout absorb the lag, and the companies using the tools get the benefit early. That is probably a good outcome for many people. It is not what is currently priced in. The market is pricing fast enterprise monetization, not just strong infrastructure demand.

[P_Rad @prashantrad]

Most investors assume the human layer shrinks as AI improves.

The actual dynamic may run the other way.

Why the gap stays structural:

Regulation

Liability

Trust

In finance, healthcare, and legal, accountability still lands on a person.

When AI gets a decision wrong, the system does not reach for the model. It reaches for the nearest human.

That creates a scaling paradox:

More AI agents = more exception moments, not fewer.

If 1,000 agent actions create 30 edge cases that need a human, 10,000 actions create 300.

The exception rate may stay flat. Human demand does not.

AI automates the standardized 80%.

The remaining 20% - exceptions, judgment calls, regulated decisions - becomes more valuable precisely because everything else has been automated.

The businesses worth owning are not fighting AI.

They are the infrastructure AI systems plug into when they hit the edge of autonomy.

[P_Rad @prashantrad]

I ran a structured head-to-head between Claude and Codex because I wanted to stop guessing. 11 tasks, same rubric applied to both: deal screens, stock research, bias checks, email drafts, code debugging. Final score: Claude 278, Codex 273 out of 340.

Five-point margin. Effective tie.

The score wasn't the useful part.

What the test revealed is that the two models don't separate by task category. They separate on one variable: whether the session has loaded context or you're starting cold.

Loaded context, active deal, live relationship, workflow already in motion: Claude. Cold research, fast first draft, code execution: Codex.

That's the decision rule.

Manager selection works the same way. Best track record means nothing if the strategy doesn't fit the mandate. The wrong allocation isn't the one with the lower Sharpe - it's the one built for a different book.

AI model selection is the same question. A model doesn't underperform because it's worse. It underperforms because you used it at the wrong point in the workflow.

The same rule applies. It was never a competition. It's an allocation.

[P_Rad @prashantrad]

#AI #DataStrategy #FinancialServices #CreditMarkets

[P_Rad @prashantrad]

The best signal in most firms isn't the data they buy. It's the data they already have and never use.

Dealer conversations in Bloomberg chat. Portfolio monitoring reports nobody read past the summary. Credit committee notes buried in email threads. Compliance logs that captured everything and surfaced nothing.

Nobody owned turning any of it into decisions.

That's the real problem. Not silos, not technology, not budget. Ownership. The incentive structure rewards finding new information. It doesn't reward extracting signal from what's already there. It's easier to justify buying new data than admitting you never used what you had. So the data sits.

The same pattern runs through every business I work with now. Call logs, intake forms, CRM entries, scheduling data. Unused because the link between that data and the next decision was never built.

The first real use case for AI in most firms isn't prediction. It's connecting the data they already have to the decisions they already make.

Most businesses already have the answer. Nobody's job is to surface it.