The analyst who learned to ship
Anant Kumar is a portfolio manager at Benefit Street Partners, a credit-focused investment firm. He's not a software engineer. But he now builds production tools that analyze 300-page credit agreements and compare loan valuations across the market. His timeline? The first version of his BDC tracker took a day. Total development time including refinements: 10–12 hours.
This isn't a demo. It's live infrastructure that changed how BSP underwrites deals. Follow-up questions from the investment committee used to take days or weeks to answer. Now Kumar says analysts "can now turn a round of comments in a day with Codex, and often answer questions during the meeting itself."
The really interesting part? BSP isn't using these tools to do more deals. They're using them to add more review capacity and rigor to underwriting within the same timeframe.
Two tools, two workflow transforms
Kumar built two systems that matter. The first is a tracker that pulls public SEC filings from business development companies to compare how different firms value the same loan. BSP can now see how its marks stack up against other holders' valuations—a competitive intelligence loop that previously required manual data gathering and spreadsheet archaeology.
The second tool is more ambitious: a credit agreement analyzer. Credit agreements routinely run 300+ pages. Investors need to identify clauses that could shift value away from lenders—provisions buried in dense legal language that can materially change a deal's risk profile. The traditional approach meant engaging outside counsel to work through the agreement page by page.
Kumar's solution? He aggregated memos from leading law firms on key clauses and risk patterns, then built an analyzer he describes as a "second brain within our organization." Upload a credit agreement, wait about 20 minutes for a first-pass analysis, then spend an hour or more on human review. BSP has since refactored the tool with GPT-4 Astra (likely GPT-4o, given the timing and OpenAI's nomenclature), and it's now running in their company-hosted portal.
The "I can build things" moment
Kumar's summary is blunt: "I've gone from being just an analyst and a portfolio manager to being a developer as well now. I can build things."
This is the wedge that matters for LLM-assisted coding. Not that professional engineers ship faster—though they do—but that domain experts can now build the tools they need without waiting for software teams to prioritize their requests.
Before Codex, creating a dashboard meant finding someone with software expertise to build a website and wire up the data. Now Kumar drafts code with AI assistance and uses it for code review. He's not becoming a full-stack engineer; he's gaining enough capability to translate his domain knowledge directly into running software.
The quality bar is also rising, not just the speed. Kumar says: "The quality of my analysis has become so much better and so much faster."
The underwriting loop
Underwriting at BSP is inherently iterative. A typical round consists of further diligence plus a revised memo responding to the investment committee's questions. Deal timelines range from a few days to a few weeks, so the number of rounds varies.
Kumar's framing here is careful: "The AI is very helpful in collating data, ingesting it, synthesizing analysis, but the ultimate investment decision needs to be made by a human." This isn't AI replacing judgment—it's AI compressing the information-gathering and synthesis steps so humans can spend more time on the actual decision.
The result is more confident decision-making, not because the AI is smarter but because "we have more information at our disposal" in the same time window. That's a subtle but important distinction. The bottleneck isn't analysis paralysis; it's access to structured information under time pressure.
What Codex actually unlocked
Kumar's closing comment is revealing: "Codex has unlocked my day, because I have many more hours in the day to do more high-level work and less routine work."
This is the classic automation promise, but the mechanism matters. He's not just offloading grunt work to an assistant. He's building durable tools that change how his team operates—tools that didn't exist before because the ROI calculation for hiring engineers to build bespoke internal finance tooling didn't pencil out.
Now the tools exist because the person who understands the workflow can build them directly.
The roadmap ahead
BSP's next ambition is connecting information across deals. Longer-term, they're eyeing continuous underwriting: monitoring loans as news breaks and examining portfolio-wide effects in real time, rather than re-underwriting only after an event forces their hand.
That's a different paradigm—moving from reactive analysis to continuous surveillance. It's also the kind of system that only makes sense if you can build and maintain it cheaply enough that the incremental value justifies the overhead.
What this tells us about AI tooling in practice
A few things stand out:
Domain experts shipping tools is the real unlock. The value isn't that Codex writes perfect code—it's that it lowers the skill floor enough that people like Kumar can translate their domain expertise into running software without going through a dev team.
Production usage looks different from demos. The BDC tracker took 10–12 hours total, not a day. The credit agreement analyzer requires an hour of human review after the 20-minute AI pass. These are real workflows with real friction, not "type a prompt, get magic" demos.
The goal isn't always more throughput. BSP explicitly chose to add rigor and review capacity rather than doing more deals. That's a maturity signal—they're optimizing for decision quality, not volume.
"Second brain" is the right metaphor. Kumar's credit agreement analyzer isn't replacing legal counsel. It's augmenting institutional memory and ensuring consistent application of known risk patterns. The human still drives, but the car has better sensors.
This case study is from OpenAI's customer showcase, so it's inherently promotional. But the specifics are concrete enough to be useful. The timelines are realistic, the workflow descriptions are plausible, and the framing around human-in-the-loop decision-making feels honest.
The real test will be whether BSP's continuous underwriting vision materializes—and whether tools built by domain experts can scale beyond their creators without becoming unmaintainable spaghetti. But for now, this is one of the clearer examples of LLM-assisted coding creating value in production finance workflows.