V

Vaishak Girish Kumar

AI / ML Researcher

{{ bootLine }}_
V Vaishak Résumé PDF
AI/ML Engineer + Researcher · Open 2027

AI systems that show their work.

Research rigor. Production instincts. Evidence over claims. I build medical VLMs and agent infrastructure for teams where wrong outputs and wasted compute are expensive.

{{ p.label }}
{{ p.value }}
Portrait of Vaishak Girish Kumar

M.S. Engineering Science (AI)

UNIVERSITY AT BUFFALO · 2027

vaishak@evidence:~
{{ l.text }}
$

SCROLL TO INSPECT ↓

Start where you are

Three ways in

The one idea

Every system here can be traced back to its cause.

One thread runs through all nine builds: an output should carry the evidence that produced it. An agent step you can replay, a retrieved chunk you can attribute a sentence to, a KV block the scheduler can account for, a finding bound to the region it came from. The clearest number so far comes from the medical work.

Inspect the projects →

FOVEA-Net · ungrounded findings per report

Published baseline~35%
FOVEA-Net target<15%

Mean IoU between a finding and its region: target above 0.70.

  1. 01

    Encode

  2. 02

    Foveate

  3. 03

    Read region

  4. 04

    Gate

  5. 05

    Report

01 — About

A model is only useful in a clinic if a radiologist can check its work.

Most of the failure I care about is not accuracy — it's confident text with nothing behind it. My research attacks that structurally: architectures where a finding cannot be emitted unless it is tied to the region that produced it, and tooling that makes the path from evidence to output visible after the fact.

That thread runs through everything here. FOVEA-Net grounds chest CT findings to 3D evidence. VectorLens traces which retrieved chunk caused which sentence. AgentTraceDAG replays an agent's context window step by step. AgentOS treats each agent's KV cache as a scheduled resource instead of a shared accident.

Before Buffalo I published two peer-reviewed papers during my B.Tech at Presidency University and spent six months on open-source NLP infrastructure at Scitonic, where the work was less glamorous and taught me more: 17 bug fixes, 1,500 lines refactored, a 30% speed-up in preprocessing.

{{ c.shown }}

{{ c.label }}

02 — Research

Research & publications

Two active lines of work: grounded generation for medical imaging, and architectures borrowed from how memory actually works.

FOVEA-Net

Active

Targeting MICCAI 2026 workshop

Read the deep dive →

A hallucination-resistant vision-language model for chest CT. Five stages take a volume to a report, and each clinical finding is bound to the 3D region that justified it — so an unsupported sentence has nowhere to come from.

  • MedNeXt-L encoder → 3D RPN foveation → region transformer → confidence-gated decoder → report aggregation
  • Target: under 15% hallucination rate against a 35% baseline, mean IoU above 0.70
  • Design grounded in a 26-paper survey of grounded report generation

SalienceFormer

Complete

Independent research · weights on HuggingFace

Read the deep dive →

A transformer with two additions borrowed from the hippocampus: a salience gate that decides what is worth encoding, and a consolidation layer that rehearses it. Same parameter budget, materially better language modelling.

  • 11.83 perplexity on WikiText-2 — 35% better than the Gemma-2B baseline
  • 15 ablation variants with significance testing, not one lucky run
  • Trained weights and configs published for replication

Peer-reviewed publications

First author · 2024

AI-DxMH: Artificial Intelligence Diagnosis for Modern Health

V. G. Kumar, M. F. Pasha, A. Prusty, D. Rajeev, G. Ganesan

A peer-reviewed system proposal for multimodal diagnostic support in resource-constrained clinics, designed around explainability and deployability rather than datacenter assumptions.

Deep dive →
Second author · 2024

On-the-fly Prompt Optimization in Multi-Agent Systems: A Comparative Study

M. F. Pasha, V. G. Kumar, A. Prusty, S. Taj

Five prompt-optimization strategies compared across multi-agent workflows. A feedback loop that refines prompts between agents at run time lifted task completion 23% on multi-step reasoning benchmarks.

Comparative study · 5 strategies

03 — Work

What ships — and what it changes

Each build shows its evidence, company use, and business or cost impact. Where a benchmark does not exist yet, I name the lever instead of inventing ROI.

All repos on GitHub
Measured / validated Built / published Targets are never presented as results

04 — Experience

Experience & toolkit

Open-Source Developer

Scitonic, via Tonic.AI · Remote

JAN – JUL 2024

Core contributor to LLM pipeline infrastructure — tokenization, preprocessing, and the data-processing paths that everything downstream depends on.

  • —17 critical bugs fixed across tokenization, preprocessing and API modules
  • —1,500+ lines of Python refactored for maintainability
  • —30% faster pandas preprocessing pipelines
  • —Regression tests authored for 10+ production modules

Graduate Researcher

University at Buffalo · Buffalo, NY

2025 – PRESENT

M.S. in Engineering Science (AI). Independent research on grounded medical VLMs and agent infrastructure alongside coursework.

  • —FOVEA-Net: grounded chest-CT report generation, MICCAI 2026 target
  • —AgentOS: per-agent KV-cache scheduling over vLLM V1
  • —Maestro: self-hosted multi-model research orchestrator, shipped
  • —26-paper survey of grounded report generation

{{ s.name }}

  • {{ i }}
    •

05 — Contact

Hiring, or want to read the draft?

I reply to everything from recruiters, engineers and researchers. If you want the FOVEA-Net writeup before it is submitted, ask — I'll send it.

Building AgentOS · per-agent KV scheduling Get in touch