ICLR 2027 · Submitted
Evaluation Collapse: When Metrics Cannot Distinguish Model Behavior
Examines when scalar evaluation metrics fail to distinguish systems with meaningfully different behavior.
Computer Science · VTU Belagavi
Prakul Hiremath
Researcher & Engineer
Hey, I’m Prakul : ) I study Computer Science at VTU, Belagavi, and spend a lot of my time thinking about how intelligent systems work, where they fail, and what we lose when uncertainty gets reduced to a single number. My research sits around information theory, geometry, and machine learning. I like starting with a question, following it wherever it leads, and building things to see if the ideas hold up. I’m looking for PhD opportunities where I can go deeper, learn from people who challenge how I think, and work on questions that matter.
I’m especially interested in what happens when uncertainty is simplified as information moves through a system, and how we can tell when that changes a model’s decisions.
01
Selected work
Research on evaluation, uncertainty, representation, and sequential systems.
NeurIPS 2026 · Poster
When Does Sequential Detection Collapse to a Scalar? A Necessary and Sufficient Characterisation
Studies when sequential detection can be represented by a scalar statistic without losing decision-relevant information.
NeurIPS 2026 · Poster
Latency-Conditioned Selection Bias in RLHF
Studies how rollout latency and asynchronous data collection can introduce selection bias in reinforcement learning from human feedback.
AISTATS 2027 · Submitted
Information Is Not Enough: A Tight, Post-Hoc Diagnostic for Representation Robustness Under Distribution Shift
A post-hoc diagnostic for assessing whether frozen representations preserve task-relevant information under anticipated distribution shifts.
arXiv · Preprint
Calibration Drift Under Reasoning: How Chain-of-Thought Budgets Induce Overconfidence in Large Language Models
Investigates how reasoning budgets relate to calibration and overconfidence in language models.
NeurIPS 2026 · E&D submission
When Confidence Metrics Fail: Sensitivity–Uncertainty Alignment as a Diagnostic for High-Risk Model Errors
Studies whether uncertainty estimates align with model sensitivity to input changes in high-risk prediction settings.
02
Selected projects
Research software and tools built around temporal systems, evaluation, and data reliability.
LOB-Latent-Regimes
Research on identifying latent microstructure regimes and early signals in high-frequency limit order book data.
Temporal Leaks
A tool for detecting look-ahead bias and accidental use of future data in time-series ML pipelines.
Latency Gym
A C++ and Gymnasium environment for exploring reinforcement-learning approaches to latency optimization in trading infrastructure.
ParallelWatch
A multivariate temporal correlation engine for anomaly detection across related infrastructure and financial metrics.
Flashback
Time-travel debugging and transformation-lineage tracking for Pandas and Polars data workflows.
HED Score
A time-aware metric designed to evaluate when a detector identifies an event, alongside detection quality.
03
Passion projects
Applied research at the intersection of biology, climate, and energy.
BIOLOOP
An algae-based approach to industrial CO₂ capture and biomass production. I contributed to the project before exiting the initiative.
S2N
A seawater-based green-hydrogen initiative exploring pathways for hydrogen production.
Project funding
₹15 lakh
MSME Idea Hackathon 4.0 — S2N seawater-based green-hydrogen project.
₹2.5 lakh
INNO MANTHAN — BIOLOOP project grant/prize.
₹1 lakh
Startup Karnataka NAIN 2.0 — BIOLOOP.
₹2 lakh
Startup Karnataka NAIN 2.0 — URJAWAVE.
These are project grants, not personal income.
Intellectual property
| Type | Application / registration no. | Title | Status / date |
|---|---|---|---|
| Patent application | 202541014688 | Integrating Neuromorphic Computing and Reinforcement Learning in Hybrid AI for Autonomous Systems | 2025 · Application |
| Patent application | 202541014687 | Innovative CO₂ Emission Filters for Vehicles: Net-Zero Carbon Solutions with Advanced Adsorption and Smart Monitoring | 2025 · Application |
| Patent application | 202541014690 | Dynamic Adjustable Sleeve D&M for Enhanced Universal CO₂ Emission Filters | 2025 · Application |
| Patent application | 202541014689 | Hyper-Adaptive Manufacturing Using Multi-Agent AI for Autonomous Production Optimization | 2025 · Application |
| Patent application | 202541032942 | EMEFF: An Explainable Multimodal Economic Forecasting Framework for Real-Time, Transparent, and Adaptive Predictions | 2025 · Application |
| Registered design | 463354-001 | Modular Algae Pod Device for Sustainable CO₂ Capture | Certificate issued 2 Jun 2026 |
| Registered design | 463353-001 | Dual-Chamber CO₂ Capture and Air Purification Device Using Algae on Biodegradable Film | Certificate issued 25 Mar 2026 |
| Registered design | 463624-001 | Disposable Algae Bioreactor Device for CO₂ Capture Using Hydrogel and Bioplastics | Certificate issued 25 Mar 2026 |
| Published patent application | 202641023750 A | A Spatial–Temporal Deep Learning Framework for Robust Human Pose Estimation in Crowded Visual Scenes | Filed 27 Feb 2026 · Published 6 Mar 2026 |
Patent applications, registered designs, and published applications are different legal categories.
04
Experience
Research across machine learning, astronomy, and applied biotechnology.
Machine learning
Capability Dynamics Researcher · ITISMU, VTU Belagavi
Working with Dr. M. N. Birje on how model capabilities emerge during training, and how the timing of detection depends on evaluation resolution. The work studies the difference between a capability first becoming detectable and becoming reliably usable across training checkpoints. Research note coming soon.
Astroinformatics
Astroinformatics Researcher · Astronomy and Calculus Club, VTU
Presented “AI-Driven, Uncertainty-Aware Anomaly Detection in Multi-wavelength AstroSat Observations” at A Decade of AstroSat Observations: Science Outcomes and Future Prospects, an international conference organised by the U. R. Rao Satellite Centre (URSC), Indian Space Research Organisation (ISRO), in Bengaluru, 30 January–1 February 2026. The conference marked ten years of AstroSat, India’s first multi-wavelength space observatory.
Bioinformatics
Bioinformatics Research · Aliens on Earth
Contributed to BIOLOOP, an AI-enabled algal biorefinery concept for monitoring and controlling algae-based biofuel and hydrogen production, and worked on S2N, a seawater-based green-hydrogen initiative. BIOLOOP was included in an AI-in-energy case study for the India AI Impact Summit 2026 and is featured on the Global AI Impact Commons. I also presented “Real-Time Monitoring of Bioaerosol Emissions in Hydrogen Reactors” at the 14th Asian Aerosol Conference (AAC 2025), Mumbai, December 2025.
Read the poster story ↗ · India AI Impact Summit energy casebook ↗
05
Education
2023–2027
B.Tech in Computer Science and Engineering
Visvesvaraya Technological University · Belagavi, Karnataka
CGPA 8.02/10
06
Get in touch
For research, collaboration, or PhD opportunities, reach out by email.