I'm a research engineer working on reliable AI systems and sequential decision-making.
I build multi-agent systems, reinforcement-learning methods, and the ML infrastructure to run them. Recent work includes a root-cause diagnosis system for distributed services at Samsung Research, a bandit-based planner for EV charging infrastructure, and audio-deepfake detection at testAIng.
I recently completed a B.Tech in Computer Science & Engineering at Vellore Institute of Technology.
Selected Work
Agentic root-cause analysis for distributed services
Samsung Research India · 2025
Designed and built a multi-agent root-cause diagnosis system end to end: specialist evidence collection, hypothesis validation, cross-incident retrieval memory, and diagnosis-only safety boundaries, using Google ADK and MCP over Kubernetes and observability infrastructure. Evaluated it on a purpose-built fault-injection testbed (OpenTelemetry Demo plus Chaos Mesh).
HOMA: multilingual agent orchestration for smart homes
Samsung PRISM · ICCCNT 2025
Led a four-person research team building a multilingual SLM-orchestrated system that decomposes multi-device smart-home commands and coordinates sequential and concurrent execution across 11+ languages, with synthetic multilingual evaluation of Gemma, Qwen and Phi-family models under edge constraints.
HERO: EV charging-station placement via bandit search
VIT Research · under review
A two-stage system for EV charging-station placement: an XGBoost city-level demand model feeds a simulator-in-the-loop multi-armed-bandit search over full station layouts. In Ann Arbor SUMO microsimulation, it outperforms facility-location, metaheuristic and bandit baselines by up to 39.9% composite reward; results are reported in a manuscript currently under review.
PrompTrend: contextual-bandit prompt recommendation
Personal project · PyPI package
A FastAPI service that recommends prompts online using contextual bandits: BERT intent classification, Redis/PostgreSQL persistence, and feedback-driven incremental regressors under stochastic exploration. Published as an installable package.
Multilingual audio deepfake detection
testAIng · 2026
Owned ML systems and research for multilingual audio-deepfake detection (ECAPA-TDNN, WavLM, mixture-of-experts), reaching 0.94 to 0.96 AUC on internal multilingual benchmarks and 1.97% EER on Hindi fine-tuning, then migrated model-heavy inference to Docker/Kubernetes with HPA, probes and Prometheus/Grafana.
ChatterBot: a scalable LLM-serving prototype
Personal project
FastAPI, Ray Serve/vLLM, Redis, RabbitMQ and PostgreSQL serving system with synchronous, queued-async and SSE streaming modes, context caching, retries and circuit breakers, and Locust load tests. Designed to scale horizontally under high concurrency.
Eldermere: a multiplayer systems project
Personal project
A Go, SvelteKit and PostgreSQL browser MUD with WebSocket room and presence state, persistent sessions, data-driven content packs, tests and Docker Compose deployment.
Selected Publications
V. Jha. “AdaptiveCareArena: Decision-Facing Evaluation of Patient World Models Under Healthcare Resource Constraints.” NeurIPS 2026 WMHS Workshop. Accepted poster; sole author.
V. Jha, A. Srivastava, B. Ashok. “Ranking Is Not Valuation: Selection-Aware Carbon Claims under Supplier Shift.” NeurIPS 2026 TCCML Workshop. Accepted.
V. Jha, B. Ashok. “BehaviorFL-MT: Federated Multi-Task Learning under Unequal Vehicle Sensors and Partial Supervision.” NeurIPS 2026 NewInML Workshop. Accepted.
V. Jha, Y. Babu S. “CA-AMFA: Context-Aware Adaptive Multi-Factor Authentication for IoT-Based Household Security Systems.” IEEE ICDSINC, 2025. Best Paper Award.
A. Jain, V. Jha, F. Alsaif, B. Ashok, I. Vairavasundaram, C. Kavitha. “Machine learning framework using on-road realtime data for battery SoC level prediction in electric two-wheelers.” Journal of Energy Storage, 97, 112884, 2024.
A. Jain, V. Jha, et al. “HOMA: Home Orchestration using Multi-Agent systems.” ICCCNT, 2025. Presented; proceedings pending.
All publications →