A $1.7 billion gap, missed entirely

AI systems that encode how your investment firm thinks.

I’m an independent consultant to investment teams: I advise where AI belongs, encode investment judgment into the workflow, and build what custom implementation requires, with training as the entry point for teams earlier in the journey. I work primarily with buy-side credit funds; the same architecture extends to equity managers and their LPs.

A generic AI model read HCA Healthcare’s bankruptcy filing and booked its lease liability at $2.2 billion. It never checked whether that claim was capped by law. It is, at roughly $500 million: a $1.7 billion gap, missed on a document it had full access to.

A practitioner’s view, not a technology sales playbook.

Send the workflow → What I do
deal_data_dictionary
What one blind test exposed15 credits · scored blind
Filed liability
HCA Healthcare books its lease liability at $2.2 billion in its FY2024 10-K
Sourced
Unaided read
Same model, no framework: carries the $2.2 billion at face value, never tests whether it is capped
Missed
Encoded framework
Applies the statute: capped at the greater of one year’s rent or 15% of the remaining lease term
Flagged
Result
Capped claim of roughly $500 million against $2.2 billion on the books
$1.7B gap
Same model · same filings · different judgment
Scored blind, 15 real credits54 vs 35 / 60
Seventeen years in credit markets
RBC Capital Markets Bank of America Deutsche Bank
In practice

What the work looks like.

Very few firms bridge investment judgment and AI output well. The single biggest lever for output quality is codifying how you actually think through a decision, so the model is guided by real expertise instead of generic prompts.

Independent benchmark
15 real credits · scored blind
The test

The same AI model, run twice on fifteen real credits: once encoded with the judgment I built over seventeen years underwriting credit, once with none. Scored blind by two independent reviewers.

What it found

A generic AI model read HCA Healthcare’s bankruptcy filing and booked its lease liability at $2.2 billion. It never checked whether the claim was capped by law. It is, at roughly $500 million: a $1.7 billion gap, missed on a document it had full access to.

The result

The encoded framework won all 15 of 15: 54 out of 60, against 35. If your team is evaluating AI for credit work, that is the question worth asking before you trust any output: what judgment is actually built in, and what has it been shown to catch.

Read the full benchmark →
Representative engagement
Buy-side credit fund · anonymized
The input

A multi-tenant, multi-document data room - 45 documents spanning executed agreements, financials, and seller materials.

What it surfaced

A conflict between the term sheet and the executed agreement, and a structural term that warranted follow-up before IC review - details the team's initial review had missed.

What changed

Seven days of manual review compressed to one hour. The investment professional still owns the judgment - this gets them there faster.

Available workflow
Portfolio exposure monitoring
The problem

An 80-name credit portfolio needs to be reviewed for concentrated sector exposure after a rapid industry re-rating - a process that currently touches every position manually.

What it surfaces

A ranked watchlist of affected positions with supporting citations - leverage ratios, maturity dates, and covenant proximity organized by severity across the book.

The result

An unstructured manual process replaced with a ranked, cited flag list - ready for analyst review.

Available workflow
Documentation and covenant monitoring
The problem

A 50-name direct lending portfolio with active covenant packages across multiple credit agreements - tracked manually in spreadsheets, changes discovered reactively.

What it surfaces

Amendment velocity, covenant proximity flags, and structural changes across the book - synthesized from raw filings and surfaced by name, not discovered by hand.

The result

Reactive, spreadsheet-driven tracking replaced with a systematic read - material changes surface before they reach the IC.

Initial screening compressed from days to hours, with a standardized extraction framework across a multi-site, multi-tenant portfolio - structured for LP reporting.

Credit fund · 2026 · anonymized

AI does not create investment judgment on its own. It magnifies what skilled managers already have.

The EigenStrategy thesis
What I Do

Practitioner judgment: delivered as training for teams getting started, and encoded into tools for teams ready to build.

Every credit professional I talk to this year opens the conversation the same way: not whether AI matters, but where to actually start, and how to tell real capability from a convincing demo. That gap, not a lack of interest, is what keeps most firms on the sidelines. Closing it is the work below.

01

AI Training for Investment Teams

For teams earlier on the AI adoption curve. Structured sessions on how AI applies to the workflows investment professionals actually live in — across credit, equity, and LP mandates. Not generic AI education for finance. A practitioner's read on where the technology creates real leverage, where it does not, and where to start. Available for individual practitioners as well as full teams.

02

Practitioner Advisory

A clear read on where AI creates real signal versus noise in your process. Workflow mapping, agent design, vendor evaluation - what to build, what to buy, what to avoid. I have sat on both sides.

03

Document Intelligence and Conflict Detection

Source hierarchy enforcement, conflict detection, and underwriting gap identification across complex deal documents. The workflow applies an encoded source hierarchy and credit review framework automatically, surfacing which source controls, what to reconcile, and what to check next.

04

Investment Process Architecture

Turning diligence, monitoring, and reporting workflows into sourced, repeatable systems that fit how investment teams already work.

Where this shows up

The workflows where domain knowledge is hardest to replicate.

iDeal screening & initial diligence
iiData room processing & extraction
iiiCovenant & indenture analysis
ivBorrowing base & collateral review
vPortfolio monitoring & reporting
viLP reporting & investor updates
viiMarket data integration & hygiene
viiiRisk assessment & stress testing
From workflow to decision

Where AI fits in the work chain, and where it does not.

The point is not to automate investment judgment. The point is to compress the manual work that prevents investment judgment from getting to the right place quickly.

Raw inputs

What the team starts with

  • Data rooms with 30 to 50 plus documents
  • Executed agreements, decks, financials, and lender materials
  • Portfolio files, filings, covenant packages, and LP reporting inputs
AI system does

Structured first-pass conversion

  • Extracts terms into a consistent frame
  • Applies source hierarchy and flags conflicts
  • Surfaces gaps, outliers, and what to check next
Human does

Judgment, escalation, decision

  • Assesses whether the flag changes the credit
  • Decides what deserves follow-up before IC
  • Uses experience to size risk and act on it
Outcome

Better workflow economics

A structured first read in hours instead of days, with analyst time shifted away from mechanical extraction and toward actual underwriting.

How engagements work

Start small, on a real problem, with no long-term obligation.

Whether your team needs to understand where AI fits before committing to anything, or is ready to build a production-grade workflow, the engagement starts where you are. The domain expertise and the architecture scale with you.

01

Training

For teams at the beginning of the journey. A structured session on how AI applies to credit workflows specifically — where it creates real leverage, where it does not, and how to think about sequencing the work. Standalone or as an entry point into a broader engagement.

02

Scoping

Where AI genuinely helps, and where it does not, including build versus buy. What data to ingest, what to pay for, and how to use it well. Security and compliance carry equal weight from day one.

03

Proof of concept

Run a POC on a live deal or workflow. You see exactly what gets built and what the output looks like - before committing to anything broader. Scoped and priced separately.

04

Implementation

Build, refine, and hand off the workflow inside how the team already operates. Documentation and a handoff your team can use and extend independently.

Writing

What I am seeing in AI, workflow design, and finance.

Short pieces on where AI is actually becoming useful in financial work, where the architecture matters more than the model, and what institutional teams are underestimating.

AI workflow architecture · 2026

The bottleneck is not access to AI

Most teams already have model access. What they do not have is a workflow architecture that converts raw information into a usable first read.

Read article →
AI in credit · March 2026

The Bloomberg terminal was built for a different world

After nine months building AI tools for credit synthesis, the hard part turns out not to be the technology.

Read article →
AI workflow · June 2026

Before the call, the context you already have

The most valuable briefing before a financial services meeting is what you already know but can't access in two minutes.

Read article →
See all writing →
About

Built from inside an investment process, by someone who lived it.

Prashant Radhakrishnan
◷
headshot
prashant_headshot.png
Prashant Radhakrishnan
Independent Consultant, EigenStrategy

I spent seventeen years in credit markets, most recently as Managing Director and Head of High Yield Trading and E-Strategy for Leveraged Finance at RBC Capital Markets. I hold a graduate degree in computational finance from Carnegie Mellon, where my interest in applying mathematics to markets took shape.

I work as an independent consultant to investment teams, engagement by engagement: scoping where AI creates real leverage, training the people who’ll use it, and implementing what makes it stick.

What has always drawn me to credit is the pattern recognition - the connections between capital structures, the signals buried in document language, the way information moves unevenly across asset classes. I have managed traders, managed risk, and deployed real capital across market cycles.

These tools are designed from the inside of an investment process. The AI extends that experience. It does not substitute for it.

Managing Director, Head of HY Trading & E-StrategyLeveraged Finance
RBC
Credit Trading
Bank of America
Credit Trading
Deutsche Bank
MS, Computational Finance
Carnegie Mellon
Questions

What teams ask before we start.

Generic tools can summarize documents. They do not reliably know which terms matter, which source controls, what conflicts need escalation, or what a credit professional checks next. That judgment lives in people who have made credit decisions with real capital. It does not exist in the foundation model itself.
A scoping conversation focused on a specific workflow or deal type, then a proof of concept on a live deal so you can evaluate the actual output before committing to anything broader. POC work is scoped and priced separately. Ongoing relationships are structured as monthly retainers once scope is defined.
Every engagement begins with a mutual NDA. I work exclusively with model configurations that do not train on client data and do not retain it beyond the engagement - a non-negotiable baseline for confidential deal materials. For teams with strict requirements, I can build inside your environment rather than routing documents through third-party platforms, scope access to the specific materials in play, and document data handling so it clears an information-security review. I have worked with pre-close investment data, names-based watchlists, and LP-facing reporting under confidentiality obligations throughout my career, and I am comfortable mapping the engagement to your existing vendor and access-control policies.
Yes. Seventeen years in public credit - high yield, IG crossover, leveraged loans - before private credit reshaped where capital flows. Private credit and direct lending are strong use cases: large data rooms, dense documents, significant manual review burden. The most interesting analytical questions often sit at the intersection of both.
Get in touch

Send the workflow. I’ll tell you if AI can actually help.

Send a workflow or data-room bottleneck. I’ll tell you whether AI can help, what the first test should be, and where it is likely to fail. Most engagements start with a single scoping conversation - no long-term obligation.

Best starting points: a data room, a covenant-monitoring process, a borrowing base review, or an LP reporting bottleneck.

Email Prashant → Follow on LinkedIn
Email
[email protected]
Web
eigenstrategy.com