On the planet of finance, the usage of AI is hard owing to privateness points, the dreaded threat of hallucinations and the guardrails. Provided that fashionable programs are more and more getting foolproof, there’s a large alternative for altering the business with brokers—all with a little bit fine-tuning.
Debarag Banerjee, chief AI and knowledge officer at L&T Finance, spoke with AIM about how the agency is steadily transferring away from rules-based automation in the direction of a future led by brokers.
Crucial and actual transformation occurring below the hood is with Challenge Cyclops, L&T Finance’s proprietary AI stack for real-time, high-accuracy credit score decisioning.
“We launched this with our two-wheeler portfolio final 12 months. It now handles 100% of these loans,” Banerjee stated. “This 12 months, we’ve prolonged it to our tractor enterprise and are getting ready to roll it out for small enterprise loans.”
Challenge Cyclops pulls collectively numerous “belief alerts”—from buyer profiles to compensation behaviour—into an ensemble mannequin that may immediately separate delinquent-risk debtors from credit-worthy ones. “You add your data and, identical to that, you get a choice,” Banerjee stated.
The way to Construct This in India?
When it comes to knowledge residency and compliance, L&T Finance is already future-proofing issues. “Even for closed-source LLMs, we insist on endpoints hosted in India to arrange for legal guidelines just like the Digital Private Information Safety Act (DPDPA),” Banerjee added.
Open-source fashions naturally present higher management. “Since we host and handle the stack ourselves, the information is safer. We guarantee contractually that our EII knowledge isn’t used for mannequin retraining.” A lot of the information used to fine-tune open-source LLMs is proprietary. Even when it’s not proprietary, the formulation below which the crew trains them makes the contextual utilization proprietary.
L&T Finance has additionally explored Indian LLMs constructed for Indic languages. “They’re begin. However the variety of parameters nonetheless issues.”
In the meantime, Challenge Cyclops, the agency’s proprietary ML stack, continues to scale. It combines fashions throughout numerous belief alerts — from buyer knowledge to compensation behaviour — and re-ensembles them to ship real-time credit score choices.
The corporate intentionally took a multi-LLM route from day one. The objective is flexibility, not being locked into any single supplier or mannequin.
“As an alternative of being tied to LLMs from anybody firm, our stack can name any mannequin our developer thinks is true for that job,” stated Banerjee. This consists of Google’s Gemini (a number of variations), OpenAI fashions by Azure, and a number of other open-source LLMs hosted on GPU-as-a-service platforms.
They’ve additionally examined Meta’s Llama household (3.1, 3.2) and efficiently fine-tuned them for efficiency similar to bigger fashions like Gemini, however with decrease inference prices.
“In one in all our different purposes, we discovered medium-sized fine-tuned Llamas performing almost in addition to a number of the premium fashions,” he famous. “We’re agnostic to geography or firm, so long as knowledge privateness is maintained and we retain full management.”
Tackling Bias and Ethics in AI Lending
When requested about moral issues in utilizing AI for credit score choices, particularly instances the place background visuals or private setting may affect mannequin behaviour, Banerjee careworn two issues: statistical validation and consent.
“Any belief sign we use has to carry up in opposition to statistically vital previous knowledge. If it’s frivolous, it will get discarded,” he stated. “We’re additionally very cautious about consent. All knowledge utilization is totally clear to the shopper.”
He acknowledged the chance of adversarial behaviour, like clients gaming the system with synthetic backgrounds. “However these sorts of patterns are caught by high quality checks and operational safeguards,” Banerjee famous. “It’s a crew effort — credit score, threat, subject ops, everybody should align for the system to work at scale.”
Conventional software program relied on infinite guidelines. With agent frameworks powered by LLMs and task-specific instruments, Banerjee stated that the hassle to construct programs has lowered drastically. “You’ll be able to create one thing with a minimalist strategy, deploy it, and let it enhance over time. These usually are not simply instruments — they’re self-improving programs,” Banerjee stated.
The way forward for agentic AI is each thrilling and inevitable for Banerjee. “Agentic AI appears to have lastly caught that proper hole,” he noticed. “We are able to already see it proving its mettle—not solely in decision-making however in emulating human features.”
He envisions brokers changing into self-improving and minimalist in design. “As an alternative of writing rule after rule, you get to a working resolution shortly, take a look at it, enhance it, and even reinforce it.”
Although we’re not totally there but, the progress is palpable. “I’ve seen demos the place brokers can do EDA, generate train-test units, construct fashions, and write documentation—all with little or no human intervention,” he stated. “That future, the place one thing like Challenge Cyclops is 90% agent-driven, can be great.”
Regulation, Black Swans & the Human Contact
Will brokers finally monitor finance and buying and selling platforms on their very own? “There are nonetheless regulatory wants—maker, checker, monitor—which is able to keep,” he stated. “And when you can create brokers for predictable failures, black swans are by nature unpredictable.”
He additionally believes AI will create jobs. “One space the place AI might generate jobs is in people enjoying each white hat and black hat—in search of methods AI can fail or be misused and determining methods to recuperate.” For L&T Finance, agentic AI is not only about tech, it’s about fixing for India’s underserved.
“I used to be there when India bought linked. Then got here digital funds. Now, the following large inflection is digital entry to credit score for the underside and center of the pyramid,” he stated. “India has the chance to leapfrog previous credit score programs as a result of its shoppers are digitally linked.”
The publish L&T Finance Hopes Challenge Cyclops Would Be 90% Agentic One Day appeared first on Analytics India Journal.
