# Research by Adith Manikonda

> Database systems, trustworthy AI, distributed systems, security and machine learning: what is being worked on, what has been published, and what is being read.

Page: https://adith.xyz/research
Author: [Adith Manikonda](https://adith.xyz/about.md) (https://adith.xyz)
Last updated: 2026-11-11

## Areas

- **Database Systems**: Explorations in analytical query execution engines, approximate query processing (AQP), adaptive sampling, and columnar storage efficiency.
- **AI Systems & Security**: Investigating LLM agent security vulnerabilities, memory trust boundaries, prompt injection resilience, and trustworthy retrieval augmented generation.
- **Distributed Systems**: Building high-concurrency event pipelines, distributed telemetry tracing, OpenTelemetry collectors, and failure resilient backend infrastructure.
- **Operating Systems & Kernel**: Analyzing TOCTOU race conditions, system call boundaries, edge compute sandboxing, and lightweight process isolation.
- **Machine Learning & Analytics**: Applying machine learning to time-series anomaly detection, industrial SCADA telemetry reduction, and computer vision medical screening.

## Active

### Memory Trust in Text-to-SQL Agents

Active Research. Stage: Experimental Evaluation & Benchmark Suite.

Evaluating state contamination, context persistence, and injection risks in conversational text-to-SQL agent pipelines.

- The problem: Multi-turn text-to-SQL agents accumulate schema context and dynamic user state in memory buffers. Adversarial or ambiguous queries cause state pollution, leading to destructive query generation or sensitive table exposures.
- Why it matters: Enterprise database copilots rely on persistent conversation memory, yet current architectures lack isolated sandboxing for schema context vs user query state.
- The approach: Building a deterministic evaluation harness that injects adversarial schema context into long-horizon text-to-SQL agent traces and measures semantic query drift and unauthorized data retrieval.
- Where it stands: Benchmark suite implemented with 400+ test cases evaluating memory pollution across multiple open-source text-to-SQL agent frameworks.
- Next: Formalizing a zero-trust memory boundary protocol for agentic database assistants.

Links: [Benchmark Code](https://github.com/Ad1th)

### Approximate Query Processing & Adaptive Sampling

Active Research. Stage: Prototype Engine & Benchmark Validation.

Sub-second analytical aggregation over multi-terabyte log streams using error-bounded adaptive sampling algorithms.

- The problem: Exact SQL aggregation over massive multi-terabyte telemetry streams incurs prohibitive latency and compute costs during exploratory backend debugging.
- Why it matters: Developers require instant confidence intervals for trace metrics rather than waiting minutes for exact full-table scans.
- The approach: Developing a dynamic stratified sampling layer integrated with columnar storage formats that continuously adjusts sample rates based on variance thresholds and query confidence boundaries.
- Where it stands: Prototype sampling engine achieving 15x-50x speedups with <2% relative error on aggregations over 50M log records.
- Next: Extending adaptive sampling to multi-table joins and sliding-window time series aggregates.

Links: [AetherQuery Repo](https://github.com/Ad1th/AetherQuery)

### TOCTOU Vulnerabilities in AI Agents

Early Research. Stage: Vulnerability Taxonomy & PoC Exploits.

Investigating Time-of-Check to Time-of-Use race conditions in autonomous file-system and API-executing AI agents.

- The problem: AI agents inspect environment state (file permission, API token, filesystem path) at time T1, then execute actions at T2 based on LLM reasoning latency. The environment state can mutate asynchronously between T1 and T2.
- Why it matters: As AI agents receive file system and shell privileges, traditional TOCTOU security gaps become amplified by multi-second LLM inference delays.
- The approach: Cataloging asynchronous state mutation patterns during agent tool calls and constructing proof-of-concept exploits that swap symlinks or permissions during agent decision windows.
- Where it stands: Identified 3 core TOCTOU vulnerability patterns in popular agent tool-calling runtimes.
- Next: Designing an atomic, transactional tool-call execution wrapper for secure agent runtimes.

### Database Query Plan Intelligence

Ongoing. Stage: Data Collection & Cost Model Training.

Machine-learning-assisted cost estimation and query execution plan optimization for non-stationary analytical workloads.

- The problem: Traditional heuristic query optimizers misestimate cardinalities when dealing with correlated predicates and skewed data distributions, resulting in suboptimal join orders.
- Why it matters: Learned query optimizers can adapt to data shifts, but require zero-overhead inference models to remain practical for production databases.
- The approach: Training lightweight tree-based ensemble models on historical execution plan traces to predict operator runtime and refine cardinality bounds dynamically.
- Where it stands: Collected 100k+ plan traces across diverse synthetic and real-world database workloads using Argus telemetry tools.
- Next: Implementing real-time plan correction during active query execution.

Links: [Argus Visualizer](https://github.com/Ad1th/Argus)

## Papers

- An Efficient Handcrafted Feature Driven Ensemble Learning Approach for Retinal Disease Screening. Published in NeLeX 2026 (IEEE Xplore), 2026. Authors: Adith Manikonda et al.. Presents a novel ensemble learning methodology combining handcrafted feature engineering with decision fusion algorithms for high-accuracy automated screening of retinal diseases in resource-constrained deployment environments.
- Wind Turbine Fault Detection using SCADA Data. Under review at IEEE Transactions on Industrial Informatics (IEEE), 2026. Authors: Adith Manikonda et al.. Introduces an anomaly detection framework for SCADA sensor telemetry from wind turbines, leveraging temporal feature extraction and dynamic thresholding to detect early component failures before catastrophic breakdown.
- Adaptive Threshold Based Time Series Data Reduction for Efficient Storage and Forecasting. Under review at Data & Knowledge Engineering (Elsevier), 2026. Authors: Adith Manikonda et al.. Proposes an adaptive thresholding algorithm for lossy compression of high-frequency time-series IoT data streams, preserving key trend inflection points while achieving up to 80% reduction in storage footprint with minimal forecasting error impact.
- Digital Twin for Smart Manufacturing. In progress, 2026. Authors: Adith Manikonda et al.. Developing a real-time digital twin system architecture combining edge sensor streams, physical system simulation, and predictive maintenance models for smart manufacturing operations.

## Patents

- [A System for Real Time Environmental Perception and Assistance for a Visually Impaired User](https://adith.xyz/patents/visual-assistance-system.md): Indian patent application 202641010249, filed 31 Jan 2026, published 13 Feb 2026.
- [Wave Energy Generator Electrical Circuit with Maximum Power Point Tracking and IoT Telemetry System](https://adith.xyz/patents/wave-energy-generator.md): Indian patent application 202641032830, filed 18 Mar 2026, published 18 Mar 2026.

## Timeline

- 2025: **Initiated Applied Systems & ML Research** (VIT Vellore / Independent Exploration). Began formal research into database query visualizers (Argus), time-series anomaly detection, and SCADA telemetry optimization.
- 2026: **IEEE Publication (NeLeX 2026)** (IEEE Xplore Conference Proceedings). Published paper 'An Efficient Handcrafted Feature Driven Ensemble Learning Approach for Retinal Disease Screening' at NeLeX 2026 conference.
- 2026: **Dual Patent Publication** (Indian Patent Office (Application Nos. 202641032830 & 202641010249)). Published patents for Wave Energy Harvester MPPT IoT Telemetry and Real-Time Environmental Perception for Visually Impaired Users.
- 2026: **Research Internship** (Indian Institute of Technology (IIT) Hyderabad). Accepted competitive research role exploring high-performance systems engineering, distributed infrastructure, and hardware-software telemetry.
- CURRENT: **Advanced Database & AI Trust Systems** (Digital Research Lab). Active research in Text-to-SQL Agent Memory Pollution, Approximate Query Processing algorithms, and TOCTOU vulnerabilities in AI agent runtimes.

## Reading

- DuckDB: an Embeddable Analytical Database, Mark Raasveldt, Hannes Mühleisen (2019). Completed. Vectorized query execution algorithms for in-process OLAP engines.
- BlinkDB: Queries with Bounded Errors and Bounded Response Times on Very Large Data, Sameer Agarwal et al. (UC Berkeley RISELab) (2013). Notes Available. Stratified sampling mechanisms for sub-second AQP error bounds.
- Neo: A Learned Query Optimizer, Ryan Marcus et al. (MIT CSAIL) (2019). Reading. Deep reinforcement learning for relational query plan generation.
- Spanner: Google's Globally-Distributed Database, James C. Corbett et al. (Google) (2012). Completed. External consistency and TrueTime API using synchronized atomic clocks.
- Dapper, a Large-Scale Distributed Systems Tracing Infrastructure, Benjamin H. Sigelman et al. (Google) (2010). Completed. Low-overhead out-of-band trace collection in microservices.
- In Search of an Understandable Consensus Algorithm (Raft), Diego Ongaro, John Ousterhout (Stanford) (2014). Completed. Leader election and replicated log management for consensus.
- SWE-bench: Can Language Models Resolve Real-World GitHub Issues?, Carlos E. Jimenez et al. (Princeton University) (2024). Notes Available. Rigorous execution-based evaluation for autonomous coding agents.
- BIRD: A Big Bench for Large-Scale Database Grounded Text-to-SQL Evaluation, Jinyang Li et al. (HKUST / Alibaba) (2023). Completed. Dirty database schemas and complex analytical query evaluation.
- TOCTOU Race Conditions in File Systems and Agent Executions, Matt Bishop et al. (2020). Reading. Asynchronous state mutation vulnerabilities during multi-step tool calls.
- Not What You’ve Signed Up For: Compromising Real-World LLM-Integrated Applications, Sahil Abdelnabi et al. (CISPA) (2023). Notes Available. Indirect prompt injection taxonomy in retrieval-connected AI workflows.
- BPF Performance Tools: Linux System and Application Observability, Brendan Gregg (2019). Completed. In-kernel tracing, system call instrumentation, and zero-copy probes.