From building Jarvis and F.R.I.D.A.Y. to developing AIOS, the BBA student is carving an unconventional path across artificial intelligence, business and decision architecture.
Most people building AI want their systems to sound confident and agreeable. Soham Kumar wants his to argue. That design choice captures a 19-year-old who has already moved from personal voice assistants to multi-agent AI architecture.
Soham is pursuing a BBA at ITM Janakpuri under GGSIPU. His formal education is in business, not computer science or engineering. Yet while many students his age are still deciding which side of technology they want to enter, Soham has been independently building the systems he wants to understand.
He began with Jarvis — an AI executive assistant with “Hey Jarvis” wake-word detection, voice input and output, multi-engine support using local models through Ollama alongside Gemini, app control, live system information and a cinematic Iron Man-inspired dashboard. Jarvis was not designed as another chat window. It was Soham’s attempt to make AI feel like an interactive operating environment.
Then came F.R.I.D.A.Y., a more technically demanding real-time voice AI project. Its architecture uses a LiveKit Agents pipeline, Deepgram for speech-to-text and text-to-speech, Gemini 2.0 Flash for reasoning, and an MCP-powered tool server connecting the assistant to web search, live news and system information. Friday pushed Soham from interface experimentation toward tool-connected voice intelligence.
For a 19-year-old BBA student, either project would have been enough to form a respectable portfolio. Soham treated them as groundwork.
His current project, AIOS, is the larger bet — a multi-agent AI system for structured business decision support. It is built around five specialized executive agents: CEO, CFO, CTO, CMO and COO. Each examines the same business problem through a defined professional domain: strategy, finance and risk, technical feasibility, market positioning, and operations.
The critical design decision is simple: the agents are not supposed to agree.
“Five AI executives nodding at each other is useless — that’s just one chatbot wearing five costumes,” Soham says. “If my CFO agent sees a financial risk that the CEO agent’s strategy ignores, I want that conflict on the table. In a real company, that argument is where the actual decision happens.”
When a CEO agent identifies a strategic opportunity and the CFO flags financial exposure, AIOS is intended to surface that contradiction rather than bury it beneath a smooth consensus answer. A separate synthesis layer compares perspectives, detects contradictions and consolidates analysis according to domain relevance.
It is an unusually specific problem for someone whose academic identity still reads “business student.” His business education is not separate from AIOS; it shapes the architecture.
Soham is also interested in valuation, private equity thinking, investment logic, HNI networks, auctions, scarcity and the perception of value. He does not present these as professional credentials. They are areas he actively studies because they converge on a broader obsession: how decisions are made when information, capital, risk and competing perspectives collide.
The hardest technical concern underneath AIOS is AI drift. As AI systems become more complex, an agent can gradually move away from its original objective or role. In multi-agent workflows, Soham is exploring a related concern he describes as “transition drift” — a working description, not an established academic term.
The idea is that every handoff between agents can reinterpret, compress or subtly alter context. After enough transitions, a capable AI system may begin solving a slightly different problem from the one the user actually asked.
His response is architectural: role anchoring, objective re-anchoring, structured context, standardized handoffs, validation against the original request, contradiction detection and an independent synthesis layer. He is exploring whether architecture can reduce drift’s probability and impact.
There is an obvious confidence behind the progression from Jarvis to Friday to AIOS. Soham is not positioning himself as a researcher with every answer or a founder with inflated numbers, funding announcements and manufactured success stories. His flex, if there is one, is considerably easier to inspect: he keeps choosing a harder system to build.
“I’ve never been interested in proving that I’m better than everyone else. I’m interested in becoming so good at what I do that the comparison eventually becomes irrelevant. While others wait for opportunities, I prefer to build something they’ll eventually have to notice.”
AIOS remains under active development. But his trajectory is becoming difficult to dismiss as casual experimentation. A business student built Jarvis, moved into real-time tool-connected voice AI with Friday, and is now confronting disagreement and drift in multi-agent systems.
Plenty of people talk about the future of AI. Soham Kumar appears more interested in having projects to point at when the conversation starts.
He is not waiting to inherit a place in technology. He is building enough that, eventually, ignoring his work may become harder than noticing it.
