AI SYSTEMS, DATA SCIENCE, AND RESEARCH
Building reliable intelligent systems.
I’m Vedant Borkute, a curious mind with a background in data science and machine learning.
Currently pursuing an MS in Computer Science at UC Riverside after earning a B.Tech in Computer Science from IIT Bombay, I work across intelligent agents, runtime security, deepfake detection, time-series forecasting, and production ML systems.
Seeking full-time roles beginning February 2027 — especially research engineer, machine learning engineer, and data scientist positions focused on intelligent agents, AI safety and security, LLM post-training, and dependable ML systems.
Now
RESEARCH INTERN: CompFly AI
Agent governance at runtime
I’m building reproducible evaluations for prompt-injection and high-stakes tool-use scenarios, improving lightweight detection pipelines, and extending OpenTelemetry traces across agent turns, tool calls, handoffs, tokens, and cost.
Read the related technical write-up →Selected work
MULTI-AGENT MEMORY: ACTIVE RESEARCH
Who gets to remember?
A research note on shared memory governance for coding-agent fleets: why the benchmark matters, what failed in earlier experiments, and the controlled write/read study we are running now.
Research note → · Real-data continual learning write-up → · ReplayGate live experiment →EE243: ADVANCED COMPUTER VISION
Concept erasure under stress
A research analysis of ESD through SPEED, with experiments on rank saturation, collateral damage, and paraphrase evasion in text-to-image diffusion models.
Read the analysis →TIME-SERIES FORECASTING
From noisy data to decisions
At Finarb AI, I built forecasting pipelines and AI-assisted analytical workflows for pharmaceutical supply-chain planning, reducing mean absolute percentage error by 90% over existing baselines.
Read the Business Wire announcement →AI FOR NAVIGATION: LLM REASONING
Comparing reasoning strategies
I tested Reactive, ReAct, Chain-of-Thought, Tree-of-Thought, and Buffer-of-Thought strategies in MiniGrid using Qwen2.5-7B-Instruct, tracking success, exploration, and token usage.
Result: Up to 55% success on LavaGapS6 and 20% on LavaCrossingS9N2.
View the project on GitHub →EE269: PAPER PRESENTATION
LLM safety & mechanistic interpretability
For EE 269 at UC Riverside with Professor Yinglun Zhu, I prepared a graduate-level paper presentation spanning four recent papers on LLM safety, mechanistic interpretability, and the bridge between them.
View the presentation on GitHub →Open-source & experiments
AGENT EVALUATION: HACKATHON
SkillFalsifier
A falsification gate for self-improving agents: generated skills must transfer to held-out and adversarial cases before they are promoted.
Try the live policy demo → · Read the technical blog → · GitHub ↗AGENT TRAINING: HACKATHON
Terminal DreamGym
A local wind tunnel for self-improving terminal agents, turning failures into practice worlds and testing recovery skills against held-out and adversarial tasks.
View on GitHub →BIG DATA: CS 226
Vegetation anomaly pipeline
A cloud-native pipeline for vegetation stress preceding Mountain Pine Beetle outbreaks in Boulder County. It combines satellite COGs, S3/PostGIS, PySpark NDVI/NDMI analysis, and Z-score detection; the report measures 15–25× indexed-query gains and about 71% recall.
View the project → Read the report →SYNTHETIC IMAGE DETECTION
Image-realiser
A CLIP-based discriminator and convolutional purifier for detecting and suppressing synthetic image artifacts, with adversarial training, residual maps, and Grad-CAM explanations.
View on GitHub →ML4SCI: SYMBOLIC REGRESSION
LM-JEPA for symbolic regression
A screening implementation for scientific discovery with a reduced-size LM-JEPA, using binary prefix notation to keep generated mathematical expressions compact and structurally valid.
View on GitHub →MULTI-AGENT MEMORY: RESEARCH MAP
Institutional memory in LLM systems
An interactive research map for shared memory, provenance, consistency, access control, and forgetting in multi-agent systems.
View on GitHub →Collaboration
- CS GSA Treasurer, UC Riverside.
- Reader/Grader for a graduate class, CS 228: Deep Learning, with Professor Zhouxing Shi — evaluated assignments and exams, gave feedback, and contributed rubric ideas.