About Research Projects GitHubSkills Experience Contact Download CV
175°
Pen Rotation
3
Training Seeds
97.2%
Target Accuracy
49
Azure AI Badges
SYSTEM ONLINE — ML & AI ENGINEER — AVAILABLE FOR OPPORTUNITIES
Bharadwaj Rachuri

Robot learning · Reinforcement learning · VLAs · LLMs.
MSc AI & Robotics (Commendation) · IJRES published researcher · 49 Azure AI certifications · UK Graduate Visa.

View Research → 📊 Live Results GitHub Profile Download CV
scroll
Bharadwaj Rachuri
IJRES
Published
MSc
Commendation
Available for opportunities · Immediate start
About

Researcher.
Engineer. Builder.

I am an MSc AI & Robotics graduate from the University of Hertfordshire (Commendation, 2025). My research trained a 24-DoF Shadow Dexterous Hand to rotate a pen using Proximal Policy Optimisation — published in IJRES in April 2025. The robotic hand you saw in the hero is built from the same architecture I trained.

Currently working independently on RL safety research (DexHackBench, TraceMonitor, Graceful Under Pressure), LLM evaluation pipelines, and on-device AI tools including Tony and CareerOS. I hold 49 Microsoft Azure AI certifications and a UK Graduate Visa with no sponsorship required.

PPO Reinforcement Learning MuJoCo Simulation LLM Engineering Azure AI Python AI Safety Sim-to-Real Transfer On-Device AI
Published Work

Peer-Reviewed Research

First-authored RL research on dexterous in-hand manipulation — published April 2025 in IJRES.

IJRES VOL.13 · ISSUE 6 · PP.164–183 · APRIL 2025

Reinforcement Learning for Robot Dexterous
In-Hand Manipulation (Shadow Hand)

Trained a 24-degree-of-freedom Shadow Dexterous Hand to perform in-hand pen rotation using PPO, phase-based reward shaping, and tactile sensor feedback. Discovered and mitigated reward hacking behaviour during training — the policy learned to exploit the reward signal rather than completing the task, then was corrected through redesigned reward decomposition. Results reproduced across 3 independent training seeds.

175°
Rotation Achieved
97.2%
Of 180° Target
18,007
Peak Reward
>90%
Touch Accuracy
3 seeds
3.118±0.047 rad
73,728
Timesteps (CPU)
Read Paper → 📊 Interactive Dashboard View Code
📄
Rachuri, B. (2025). Reinforcement Learning for Robot Dexterous In-Hand Manipulation of Objects (Shadow Hand). IJRES, Vol. 13, Issue 6, pp. 164–183. ISSN: 2320-9364. ijres.org ↗
Built Work

Projects

Real projects with real results — every metric on this page came from an actual experiment, training run, or test log. Click any entry to explore.

Latest — Robot Learning from Human Video
🤖
OPEN SOURCE · OCT 2026
phone-to-panda — Phone Video of a Hand to a Robot Policy
14 phone clips of a hand, retargeted to a Panda arm in the LIBERO simulator and stored relative to the object, so one clip replays on any layout. ACT trained and SmolVLA fine-tuned on the resulting demonstrations, plus a study of how far the motion can be sped up.
→ 93% replay success on unseen layouts · 101 demonstrations · ACT and SmolVLA 58% from camera images
LeRobotLIBEROSmolVLAACTMediaPipe
View on GitHub →
Published Research
📄
PUBLISHED · IJRES
Dexterous In-Hand Manipulation — Shadow Hand PPO
Peer-reviewed research on training a 24-DoF Shadow Hand using PPO. Phase-based reward shaping, tactile sensor integration, reward hacking discovery and mitigation.
→ 175° rotation · 97.2% of target · 3.118 ± 0.047 rad
PPOMuJoCoPyTorchSB3Published Research
Click to explore →
Completed — Live Results
📊
LIVE TOOL
Shadow Hand Reward Hacking Visualiser
Interactive dashboard on the published research data. Reward hacking emergence, KL divergence, touch sensor accuracy across all 5 fingers, seed reproducibility.
→ 7 interactive charts · 92 training episodes · Real paper data
Chart.jsJavaScriptAI Safety
Click to explore →
🔍
LIVE PROJECT
LLM Evaluation Workflows — Hallucination & Safety
End-to-end LLM evaluation engine — hallucination detection, safety guardrails, jailbreak detection, RAG pipeline. Zero external dependencies, fully CPU.
→ 14 PASS · 6 WARN · 0 FAIL · Avg score 0.847
PythonRAGLLM Safety
Click to explore →
Completed — Independent Builds
🧠
INDEPENDENT
On-Premise SLM Document Intelligence
Privacy-preserving local document Q&A. Phi-3 Mini via Ollama + ChromaDB RAG + Flask REST API + SQLite logging. All inference local — zero data egress.
→ 7 working endpoints · Avg groundedness 0.834 · 100% local
Phi-3 MiniOllamaChromaDBFlask
Click to explore →
🔬
INDEPENDENT
Vision-Based Defect Inspection REST API
MobileNetV2 transfer-learning CNN classifies product images as defect/no-defect. Flask REST API, Dockerised, AWS-deployment-ready. Built and tested locally.
→ 98.3% train · 91.7% val · 26/26 tests passing
TensorFlowMobileNetV2DockerAWS
Click to explore →
🚗
100 / 100
Autonomous Robotic Navigation — Unity
Sensor-driven autonomous vehicle in Unity Engine — 7-sensor raycast array, lane-keeping, real-time obstacle avoidance. Scored 100/100 in MSc assessment.
→ Full autonomous navigation · 7-sensor array · MSc 100/100
UnityC#Sensor Fusion
Personal AI Tools — Active Development
🤖
PERSONAL TOOL · MAY 2025 – PRESENT
Tony — Local JARVIS-Style AI Assistant
Local-first AI assistant with a skill-based context loader. Each skill is a SKILL.md folder — when your message matches triggers, that skill loads into context. Multi-LLM backend (Ollama), markdown vault, autonomy feed, security architecture layer. Python + Flask + HTML.
→ Skill-based context · Multi-LLM · Security layer · Active development
PythonFlaskOllamaLLMActive
Click to explore →
🎯
PERSONAL TOOL · MAY 2025 – PRESENT
CareerOS — AI Career Operating System
13-department local-first AI system for job hunting — Adzuna API, ghost-job detection, ATS scoring, resume writer, Gmail monitor, analytics. Human approval gate — never auto-submits. LangGraph orchestration in progress.
→ 13 departments · Adzuna API · ATS · Resume factory · Active
PythonStreamlitSQLiteLangGraph
Click to explore →
Upcoming Research
📉
TESTING NOW
Graceful Under Pressure
First systematic tactile-sensor failure robustness study on 24-DoF in-hand RL. Ablates dropout, stuck-at faults, calibration drift, and additive noise on the trained Shadow Hand PPO.
→ Active testing · Target: IEEE RA-L / Frontiers Robotics
MuJoCoPPOSim-to-Real
⚡
IN DEVELOPMENT
DexHackBench
First reward-hacking benchmark and trace-based detector for dexterous in-hand manipulation. KL divergence and correlation-decay detectors on phase-based reward shaping. Confirmed novel gap vs. arXiv:2507.05619.
→ Ongoing · Target: IEEE RA-L / NeurIPS Datasets & Benchmarks
PPOMuJoCoReward Hacking
🛡️
IN DEVELOPMENT
TraceMonitor
CPU-resident SLM runtime safety monitor for reward hacking in dexterous RL. Phi-3 Mini reads textualized proprioceptive/reward traces — first SLM monitor for in-hand manipulation.
→ Ongoing · Phi-3 Mini · Target: IEEE RA-L / IROS / NeurIPS Safety
Phi-3 MiniOllamaRL Safety
Live on GitHub

Source Code

Fetched live from GitHub — real repositories, real commit history.

FETCHING REPOSITORIES...
View All on GitHub →
Technical Toolkit

Skills

Reinforcement Learning
PPOSACMuJoCoSB3Reward ShapingSim-to-RealGymnasium
Deep Learning
PyTorchTensorFlowKerasCNNsTransformersTransfer Learning
LLMs & Generative AI
RAG PipelinesPrompt EngineeringChromaDBOllamaFine-tuningRLHFAI Safety
MLOps & Cloud
Azure AI FoundryMLflowFabricAWSDockerCI/CDREST APIs
Programming
PythonC#JavaSQLRFlaskFastAPI
Robotics
Shadow HandMuJoCoUnity EngineSensor FusionAutonomous Systems
49

Microsoft Azure AI & Fabric Badges

Comprehensive certification across the full Azure AI platform — foundations to responsible AI deployment.

LLMsRAGFine-tuningPrompt FlowResponsible AINLPComputer VisionMLflowAI FoundryFabric
Timeline

Experience

May 2025  →  Present
Machine Learning & AI Engineer
Independent / Swayam Ltd · London, UK · Remote
Designing, developing, and evaluating ML and deep learning models. Contributing to AI systems incorporating LLMs, SLMs, and Vision Models for client-facing projects. Building RL pipelines using PPO and MuJoCo, sim-to-real transfer workflows, SQL data pipelines. RESTful APIs with Flask, Azure AI and AWS integration, CI/CD pipelines maintained throughout.
PythonPyTorchTensorFlowLLMs/SLMsPPO/MuJoCoFlaskAzure AIAWS
Get In Touch

Let's Talk

ML & AI Engineer.
Available across the UK.

Open to Machine Learning, AI Engineering, Data Science, and Research Engineer roles across the UK. Graduate Visa — no sponsorship required. Available immediately.

✉ bharadwaj.r.career@gmail.com 📞 07586 362964 in linkedin.com/in/bharadwaj-rachuri ⌥ github.com/br23aay 🌐 br23aay.github.io
✅ Message sent! I'll get back to you within 24 hours.
Something went wrong — email me directly at bharadwaj.r.career@gmail.com