Sakhinala Sanjay Bhargav/AI Engineer/Hyderabad

I build AI systems that hold up outside the notebook.

I work across generative AI, applied AI and machine learning. Recent work: IndicRAG, an agentic RAG system that answers questions over research papers in English and 11 Indian languages; AgentOps, tracing and evaluation for LLM agents; and a reinforcement-learning optimiser that needs 72% fewer simulations to design a 6G antenna. Along the way: three IEEE papers and a filed patent.

See the work sanjaysakhinala@gmail.com

Open to entry-level Generative AI, Applied AI and AI/ML engineering roles.

Portrait of Sanjay Bhargav
Sanjay Bhargav
B.Tech ECE, Bapatla Engineering College, 2025
12
languages in IndicRAG — English plus 11 Indian languages
72%
fewer CST simulations than GA/PSO baselines (420 vs ~1,500)
3
IEEE conference papers — one published, two in press
1
Indian patent filed, hexagonal patch antenna at 28 GHz
Selected work

Generative AI, applied AI, and the research behind them

Two projects in depth — an LLM agent system and an applied RL system — then eight more, applied-AI work first. For each: the problem, what I built, and what it measurably changed.

IndicRAG architecture: hybrid retrieval feeding a LangGraph agent with a reflexion loop
Fig. 1 — v2 replaced a single retrieval pass with an agent that can decide it needs another source.
Generative AI · Open source v2.6

IndicRAG — document QA for Indian languages

Problem
Retrieval tooling assumes English. Questions and documents in Telugu, Hindi, Tamil and others get poor matches and confident wrong answers.
Approach
BAAI/bge-m3 dense vectors fused with BM25 through Reciprocal Rank Fusion, reranked by bge-reranker-v2-m3. The v2 rewrite turned the pipeline into a six-node LangGraph agent with six tools and a reflexion step that scores each draft for faithfulness and completeness before answering.
Result
12 languages, live search across arXiv, Semantic Scholar and OpenAlex alongside your own corpus, streamed answers, and failover across any OpenAI-compatible LLM provider. Now at v2.6, which adds index reconciliation and backups.

LangGraph · HuggingFace · ChromaDB · NLLB-200 · Gemini API · FastAPI

Why I rewrote it as an agent Source on GitHub
Five-stage pipeline: design sampling, CST full-wave simulation, a blended surrogate with uncertainty, a SAC agent, and the final optimised antenna
Fig. 2 — The agent explores on a fast surrogate and only pays for a full CST simulation when the surrogate is unsure.
Applied AI · Research Accepted, WAMS 2026

Uncertainty-aware RL for 6G antenna design

Problem
Tuning an antenna with genetic algorithms or particle swarms took roughly 1,500 full-wave CST simulations per run, each taking hours.
Approach
A Soft Actor-Critic agent in a custom Gymnasium environment, trained against a blended LightGBM / MLP / RidgeCV surrogate. Its uncertainty estimate decides when a real simulation is worth running, and the reward only accepts geometries that can actually be fabricated.
Result
~1,500 → 420 simulations
72% fewer CST calls, 40% fewer episodes to converge, and a final design at −52.2 dB return loss.

PyTorch · Stable-Baselines3 · LightGBM · Optuna · Gymnasium · CST API

Read the write-up

Also built

  • Meridian

    A reference prior-authorisation workflow for healthcare payers. Every AI finding must cite an exact document span or policy clause or it is rejected, low-confidence findings go to a human, and no denial can be issued without a recorded attestation. Synthetic data only.

    Python · LLMs · citation grounding · human-in-the-loop · state machine

    Code

  • NipunyaMatch

    A recruitment assistant: upload resumes and a job description, get an auditable ranked shortlist, then ask “why is A above B?” and get an answer that cites specific skills.

    FastAPI · Streamlit · Gemini · OpenRouter · SQLite · embeddings

    Live demo Code

  • AgentOps

    A small LangSmith of my own: records agent runs as execution graphs, scores answers for faithfulness and hallucination, A/B-tests prompt versions, and tracks cost and latency per run.

    FastAPI · Next.js · TypeScript · PostgreSQL · React Flow · Python SDK

    Code Live demo

  • VocalForensics

    Evidence reports for vocal tracks, scored per cue and reviewed by a human. It deliberately never answers “is this AI?”, because a wrong verdict about a named artist does real harm; it reports what it found and what it couldn’t measure.

    Python · signal processing · audio fingerprinting · PyTorch

    Code

  • AI Data Scientist

    Seven CrewAI agents that take one dataset from loading and cleaning through modelling to a written report. The LLM orchestrates; deterministic pandas, scikit-learn and scipy tools do the maths, so results are reproducible.

    CrewAI · Gemini / OpenAI / Anthropic · scikit-learn · pandas · Docker

    Code

  • CipherChat

    An end-to-end encrypted messenger on the Signal Protocol (X3DH and Double Ratchet) with per-message forward secrecy, TLS transport, and trust-on-first-use key pinning.

    Python · X25519 · AES-256-GCM · TLS · TCP

    Code Write-up (v1)

  • EuroSAT land use

    A CNN for satellite imagery, 89% accuracy across ten land-use classes, with Grad-CAM maps to show what the model is looking at.

    PyTorch · TorchVision · OpenCV · Grad-CAM

    Code

  • ResolverLab

    Benchmarks DNS resolvers across 50+ providers for latency, block rate and cache behaviour, with percentile statistics instead of single averages.

    Python · network programming · statistics · CLI

    Write-up Code
Research

Papers and a patent

Antenna design for 6G came first; the RL work grew out of how slow it was to optimise those antennas by hand.

  1. [1]

    Uncertainty-Aware Reinforcement Learning System with Blended Surrogate Models for Electromagnetic Structure Optimization

    WAMS 2026, BVRIT Hyderabad.

    Accepted
  2. [2]

    Bandwidth Optimization of Slotted Circular Patch Antenna for 6G Ultra-Fast Data Transfer and Brain-Computer Interface Application

    16th ICCCNT 2025, IIT Indore.

    Proceedings submitted to IEEE Xplore

    In press
  3. [3]

    Bandwidth Enhancement of Slotted Hexagonal Patch Antenna for 6G Ultra-Fast Data Transfer and Brain-Computer Interface

    ICMOCE 2025, IIT Bhubaneswar.

    doi:10.1109/ICMOCE64100.2025.11076991

    Published
    17 Jul 2025
  4. [P1]

    Design of Hexagonal Patch Antenna at 28 GHz

    Indian patent application No. 202541014595 A. Inventor: Sakhinala Sanjay Bhargav.

    Published
    7 Mar 2025
About

Electronics engineer turned AI engineer

I trained as an electronics and communication engineer, and my first research was on antennas. Optimising them meant waiting hours for each simulation, which is how I ended up teaching a reinforcement-learning agent to decide which simulations were worth running at all.

Since then most of my work has been about getting models to behave in the real world: retrieval that works in Indian languages, agents that check their own answers, and the tracing and evaluation tooling needed to tell whether any of it is improving.

I'm based in Hyderabad and looking for an entry-level role in generative AI, applied AI or AI/ML engineering, on a team that ships.

Training

Structured programmes

Four training programmes between 2024 and 2025, newest first.

Generative AI

SkillHive Connect · Aug–Dec 2025
  • LLMs, transformers, LangChain and retrieval-augmented generation in Python
  • Fine-tuned LLaMA-family models
  • Built chatbots, a semantic search engine, a YouTube transcript summariser and a voice-based medical assistant

AI/ML

Google · Jan–Mar 2025
  • Computer-vision modules in Python and OpenCV for image detection in mobile apps
  • Evaluated and deployed ML pipelines on Vertex AI
  • Wired real-time inference into end-to-end app workflows

Cybersecurity

Palo Alto Networks · Oct–Dec 2024
  • Firewall policy tuning to tighten access control and cut false-positive alerts
  • Threat-log analysis and SIEM correlation workflows
  • Secure network design, intrusion detection and endpoint protection basics

Applied data science

Altair RapidMiner · Apr–Jun 2024
  • Preprocessing, feature engineering and evaluation pipelines in RapidMiner
  • Benchmarked SVM and Random Forest classifiers
  • Exploratory data analysis and reporting
Toolkit

What I work with

Generative AI & LLMs
RAG and hybrid retrieval, agentic workflows (LangGraph, CrewAI, LangChain), tool use, reflexion, LLaMA fine-tuning, prompt engineering, Gemini and multi-provider LLM APIs
LLM evaluation & ops
Faithfulness and hallucination scoring, tracing, cost and latency tracking, provider failover
Retrieval
ChromaDB, FAISS, BM25, dense embeddings (bge-m3), cross-encoder reranking, NLLB-200 translation
Machine learning
Deep learning, reinforcement learning (SAC, PPO), transformers, computer vision, multilingual NLP, surrogate modelling
Frameworks
PyTorch, TensorFlow, HuggingFace, scikit-learn, Stable-Baselines3, TorchVision, OpenCV, LightGBM, Optuna
Shipping
FastAPI, Docker, Redis, MLflow, Vertex AI, Google Cloud, Streamlit, REST APIs, CI/CD, Git
Data & systems
Python, pandas, NumPy, SQL, Linux, shell, TCP/IP, DNS, applied cryptography
Writing

Notes from the work

Longer explanations of the projects above — the decisions, the dead ends, and what I'd do differently.

All writing

Contact

Let's talk

Hiring for a GenAI, applied AI or ML role — or have a problem that needs one solved?

Email is the quickest way to reach me. The form works too — it lands in the same inbox.

What's it about?
Who's writing?
Your message

A line on the role and the team is plenty. A link to the job description helps.

Goes straight to my inbox · Ctrl+Enter to send