|
🔬 Research with receipts POPE audit, public code, dataset, PyPI package, and Zenodo archive. |
🖥️ Systems that measure Cloud-edge GPU execution, offline rescue mesh, and evidence-aware RAG. |
🛠️ Upstream engineering Submitted patches across LangChain, Future-AGI, EleutherAI, and Hugging Face. |
class KesavKumarJ:
name = "Kesav Kumar Jayakumar"
role = "IT Graduate | First-Author AI Researcher"
education = "B.Tech Information Technology | CGPA 8.41 / 10"
institution = "Sri Krishna College of Technology"
focus = [
"Vision-language model evaluation",
"Distributed cloud-edge GPU systems",
"Reliable, reproducible AI engineering"
]
2026_highlights = [
"Published VisualPC at ICCET 2026",
"POPE audit under review at NeurIPS VLM4RWD",
"National Finalist, Snapdragon Multiverse Hackathon",
"Runner-Up, IBM SkillsBuild international challenge"
]
motto = "Measure first. Build right. Ship with evidence."Languages
AI, Research, and Frameworks
Cloud, Data, and Quality
Each badge links to a public project or reproducibility artifact. Competition figures and academic results below are resume-sourced where no official public result page is available.
These public cards provide a compact code snapshot. The evidence wall above remains the source of truth for verified project outcomes.
🔬 Token-Set Choice Confounds POPE | Under review, NeurIPS 2026 VLM4RWD workshop
Token-Set Choice Confounds POPE: A Systematic Audit of Yes/No Extraction in VLM Hallucination Evaluation is under review for the NeurIPS 2026 Workshop on Grounded and Faithful Vision-Language Models, VLM4RWD, Sydney.
- Traced a tokenizer prefix-space artifact that deflated POPE hallucination scores by up to 7 F1. Correcting the logit read-out raised the LLaVA-1.5 baseline by 6.13 points to 82.21 F1.
- Re-evaluated 6 published mitigation methods and found reported gains shrink or reverse against the corrected baseline.
- Confirmed the confound across 4 VLMs, 3 tokenizer families, 9,000 queries, and 90 GPU-hours. Attention probing measured Cohen's d = -0.46.
- Released
pope-auditversion 0.1.1 on PyPI and 9,000 prediction records on Hugging Face for reproducibility. Source code is archived on Zenodo with a DOI.
🖥️ VisualPC | Published at ICCET 2026, Chennai
VisualPC: A Lightweight and Measurable Framework for Experimental Evaluation of Distributed GPU Execution was published at ICCET 2026, Chennai, in March 2026. ISBN: 978-81-986418-1-6.
- Designed a hybrid cloud-edge Platform-as-a-Service. A FastAPI priority scheduler routes CUDA workloads through a Tailscale VPN mesh to cloud GPUs and Raspberry Pi 4B edge nodes.
- Built a Next.js dashboard that streams live worker telemetry through Server-Sent Events.
- Measured queue wait, execution, and network components across tensor payload sizes, with pytest integration tests, Playwright end-to-end tests, JWT authentication, and Dockerized deployment.
| Project | Patch summary | Status |
|---|---|---|
LangChain, langchain-openai |
Structured-output streaming warning fix extending a maintainer change to sync and async generators. Validated against 517 tests, Ruff, and mypy. +8 / -2 |
PR #38625, closed |
Future-AGI, agentcc-gateway in Go |
Founder-invited MCP schema-validation patch that prevents empty tool arguments from skipping required-field checks. Includes a regression guard. +43 / -1 |
PR #1405, open |
EleutherAI, lm-evaluation-harness |
Just-in-time native-Python @property lazy loader replaces eager dataset loading in lm_eval/api/task.py, reducing multi-task startup memory with no new dependencies. +49 / -7 |
PR #3862, open |
Hugging Face, lighteval |
Fail-fast static linter for custom evaluation tasks using inspect and typing, refactored through Ruff C901 with an in-memory pytest suite. +220 |
PR #1268, open |
| Project | Stack | Evidence-rich outcome |
|---|---|---|
| 🧭 RepoMind | FastAPI · React · GPT-5.6 Terra · MCP | Repository preflight engine that orchestrates four parallel GPT-5.6 processes restricted to read-only workspace tools. An application-layer firewall drops output that fails strict file-path and line-range validation. Shipped with a React dashboard, CLI, and local stdio MCP server. |
| 📡 Sankat-Mochan | Kotlin BLE mesh · LoRa · FastAPI · On-device LLM | National-finalist off-grid SOS network: Android phones relay compressed voice reports over Bluetooth Low Energy, then bridge them through a 433 MHz LoRa gateway on Raspberry Pi and Arduino. Base model · GGUF |
| ⚽ DecisionLens | RAG · IBM Granite 3.1 · BM25 + embeddings · React | IBM SkillsBuild runner-up system for football VAR calls. A hybrid retriever over 593 rule chunks feeds Granite 3.1 8B under a strict JSON schema. It abstains when evidence is weak and scored full marks on citation, keyword, and decision-type checks in a 50-question adversarial suite. |
| 🌿 YOLOv10 + CBAM Leaf-Disease Detection | PyTorch · TensorRT | Faculty-invited thesis: extended a YOLOv10-Nano, 96.2% mAP@0.5, and CBAM-MobileNetV2, 93.6% accuracy, pipeline to 5,010 images. INT8 TensorRT layer fusion reached sub-100 ms Jetson Nano latency under 50 MB. |
- National Finalist, Snapdragon Multiverse Hackathon, Bengaluru, July 2026, for Sankat-Mochan.
- Runner-Up, 2nd place, IBM SkillsBuild "AI Inside the Match" International Challenge, June 2026. Awarded $1,250 among 3,381 registrants and 307 projects, judged by a 19-member IBM panel.
- Research Collaborator, MSME Government Grant, 2025. Faculty-invited to design machine-learning algorithms for a sanctioned social-media fake-ID-detection project worth INR 13.5L, about US$14,000.
⚙️ DevOps Intern | iStudio | June 2025 to August 2025
MongoDBDockerJenkinsGitHub Actions
- Automated GitHub Actions deployments, cutting release cycles by 25% and removing manual steps.
- Built live MongoDB analytics dashboards for demos and ran Docker and CI/CD reliability checks that held 97.8% test-phase uptime.
☕ Java Development Intern | NxtLogic Software Solutions | April 2025 to May 2025
JavaSpring BootAWS EC2
- Integrated analytics modules into Java REST APIs and tuned Spring Boot with EC2 deployment strategies.
- Improved backend response time by 18% and cut server error rates by 12% under sustained load testing.
| Programme | Institution | Result |
|---|---|---|
| B.Tech, Information Technology October 2022 to April 2026 |
Sri Krishna College of Technology, India Autonomous, NAAC A, NBA-accredited |
CGPA: 8.41 / 10 10/10, O, Outstanding in the Semester 8 VisualPC capstone Cumulative top-5% department cohort rank |
- Relevant coursework: Artificial Intelligence, Data Visualization, Computer Networks, Cloud Computing and Containerization, Data Structures and Algorithms, and Operating Systems.
- NVIDIA DLI, Deep Learning and AI on Jetson Nano, 2024.
- IEEE Student Member and IEEE Computer Society Member.
AI and ML PyTorch · Hugging Face · QLoRA/PEFT · scikit-learn · RAG · LLM evaluation · NF4/INT8 · TensorRT
Cloud and DevOps AWS EC2/S3/VPC · GCP · Docker · GitHub Actions · Jenkins · Linux · Tailscale
Web and Data React · Next.js · Spring Boot · FastAPI · REST · PostgreSQL · MySQL · MongoDB
Testing pytest · Playwright · JMeter · Selenium · Postman · Git · Ruff · mypy
