I am an AI Engineer and Researcher working on systems that reason, retrieve, remember, evaluate, and adapt.
My work currently sits across:
LLM Systems → retrieval, agents, evaluation, post-training
Memory → persistent context, retrieval, agent memory
Language Models → multilingual + low-resource NLP
AI Systems → inference, optimization, reliability
Research → evaluation, fairness, retrieval stability
Physical AI → robotics, embodied agents, perception → action
I am especially interested in where these areas begin to converge:
The question behind much of my work is simple:
How do we know an intelligent system actually became better — rather than simply becoming more complex?
That is why I care deeply about evaluation, reproducibility, failure analysis, and measurable system behavior.
|
I am interested in AI systems that can accumulate useful experience instead of starting nearly from zero every interaction. Areas I want to explore:
Core question
|
I am increasingly exploring intelligence beyond text-only environments. Areas I am interested in:
Long-term direction
|
An evaluation-first RAG platform designed to detect whether changes to retrieval or generation actually improve the system.
Documents
│
▼
┌───────────────┐
│ Retrieval │
│ 3 strategies │
└───────┬───────┘
│
▼
┌───────────────┐
│ LLM │
│ 3 providers │
└───────┬───────┘
│
▼
┌─────────────────────────────┐
│ Evaluation │
│ │
│ Retrieval Hit Rate │
│ Faithfulness │
│ Answer Quality │
│ Latency │
│ Regression Detection │
└──────────────┬──────────────┘
│
▼
GitHub Actions Gate
Current evaluation
100% retrieval hit rate · 92% faithfulness
Changes to prompts, retrieval configurations, or model behavior are evaluated through CI regression gates before being accepted.
Treating RAG quality as a software regression problem.
A four-agent asynchronous research system operating across US, NSE, and BSE markets.
┌────────────────┐
│ User Query │
└───────┬────────┘
│
┌────────────────┼────────────────┐
│ │ │
▼ ▼ ▼
Research Agent Sentiment Agent Market Agent
│ │ │
└───────────┬────┴────────────────┘
▼
Analysis Agent
│
▼
LLM Evaluation
│
▼
Portfolio Optimizer
Combines:
- asynchronous agent orchestration
- FinBERT sentiment
- structured reasoning
- LLM-as-judge evaluation
- Markowitz optimization
The research problem I care about here is not simply:
“Can multiple agents talk?”
It is:
How do agent state, memory, handoffs, evaluation, and failure propagation affect the reliability of the final system?
A Python library and CLI for adapting pretrained language models to new languages and domains.
Tokenizer
↓
Vocabulary Adaptation
↓
Embedding Initialization
↓
Continued Pretraining
↓
Supervised Fine-Tuning
↓
DPO
Supports:
Tokenizer Expansion · Vocabulary Replacement · Embedding Adaptation · Weight Initialization · Continued Pretraining · SFT · DPO
The broader research question is how tokenization and representation capacity influence adaptation, especially for languages poorly represented during pretraining.
A GPU inference study measuring the effect of TensorRT compilation on ResNet-50 serving performance.
| Result | |
|---|---|
| Throughput | 97 → 347 FPS |
| Throughput gain | 3.55× |
| P99 latency | 4.3× lower |
| Accuracy loss | 0 |
Stack:
PyTorch → ONNX → TensorRT 11 → FP16 → CUDA 12.4
The focus is not only optimization.
It is understanding where performance improvements come from and measuring them reproducibly.
Models and experiments across:
తెలుగు Telugu · தமிழ் Tamil · ಕನ್ನಡ Kannada · മലയാളം Malayalam
I study whether dedicated language models can outperform multilingual models when data and compute are constrained.
Evaluation spans:
Language Modeling
├── Loss
├── Perplexity
└── Bits-per-byte
Tokenization
└── Tokenizer efficiency
Downstream Transfer
├── Sentiment
└── Named Entity Recognition
Compared against multilingual baselines including mGPT, mBERT, and XLM-R.
A systems study of reward-model inference for RLHF pipelines.
PyTorch Eager
│
│ benchmark
▼
ONNX Runtime
│
▼
Execution / Graph Analysis
The key finding:
The observed CPU speedup is primarily explained by ONNX Runtime execution and graph optimizations, rather than C++ itself.
Research intersection:
RLHF Infrastructure × Inference × Benchmarking × Systems Analysis
A longitudinal evaluation framework for studying whether model fairness remains stable across releases.
| Scope | Scale |
|---|---|
| Model versions | 12 |
| Providers | 5 |
| Model families | 6 |
| Bias benchmarks | 7 |
| Evaluation examples | ~200K |
The central idea:
Fairness should be regression-tested across model releases, not measured once and forgotten.
Status: Under double-blind review.
Researching certifiable top-k retrieval stability under multi-agent communication.
Datasets:
HotpotQA · 2WikiMultihopQA · MuSiQue
The work studies whether retrieval rankings can remain stable under bounded perturbations introduced during agent communication.
Original Query
│
▼
Retriever ──────────────► Top-k
▲
│
Agent Communication
│
▼
Perturbed Query
│
▼
Retriever ──────────────► Top-k'
│
▼
Stability Certificate
Rather than asking only:
Did retrieval change?
I am interested in:
Can we determine when retrieval is guaranteed not to change?
This is one of the directions I want to push much deeper.
Modern agents can reason over impressive context lengths, but context is not memory.
I am interested in architectures where an agent develops persistent internal state across interactions:
┌─────────────┐
│ Working │
│ Memory │
└──────┬──────┘
│
┌────────────┼────────────┐
▼ ▼ ▼
Episodic Semantic External
Memory Memory Retrieval
│ │ │
└────────────┼────────────┘
▼
Agent Reasoning
│
▼
Action
│
▼
Experience
│
└──────────► Memory Update
Questions I want to investigate:
- What deserves to become long-term memory?
- When should memory be retrieved?
- How should memories decay?
- How should contradictory memories be resolved?
- Can agents consolidate multiple experiences into abstractions?
- How should agent memory be benchmarked?
- How should memory operate in multi-agent systems?
- How does memory change long-horizon reasoning?
The direction I find most exciting is moving intelligence from:
Prompt → Model → Response
toward:
Perception
↓
World Understanding
↓
Memory
↓
Reasoning
↓
Planning
↓
Action
↓
Environment
└──────────► New Experience
I am especially interested in:
My long-term research interest is in systems that can perceive, reason, remember previous interactions, plan, and act in physical environments.
I like working across the full loop:
For me, a model benchmark is rarely the end of the project.
I want to know:
Why did it improve?
Where does it fail?
Does the improvement generalize?
What changed internally?
What happens under distribution shift?
Can the result be reproduced?
Will the system remain reliable after the next change?
Transformers · Sentence Transformers · SentencePiece
CPT · SFT · DPO
RAG · Agents · Evaluation · LLM-as-Judge
Retrieval · Structured Generation · Memory
MLflow · Prometheus · Evidently
Intelligent Systems
│
┌────────────────┼────────────────┐
│ │ │
▼ ▼ ▼
Models Memory Agents
│ │ │
└────────────────┼────────────────┘
▼
Reasoning
│
┌────────────┴────────────┐
▼ ▼
Perception Planning
│ │
└────────────┬────────────┘
▼
Action
│
▼
Physical World
I want to work toward AI systems that don't just produce better responses.
I want to understand how to build systems that can:
reason · remember · adapt · learn · perceive · plan · act

