// AI systems · edge vision · rover autonomy

I build intelligent systems that survive deployment, users and rough terrain.

CS (AI and ML) undergrad at VIT Chennai. I build RAG pipelines
from semantic retrieval backends to sub-40 ms vision inference on Jetson-class edge hardware.

AI / MLComputer visionAutonomous systemsMars rover engEmbeddedFull-stackResearch

SIH 2025Grand Finalist
AIR 1DD Robocon, Round 1
20th worldwideIntl. Rover Challenge 2025
2 papersunder peer review
Ranbir working with an angle grinder in a robotics workshop
UNIT RK-26 AI · CV · ROBOTICS
Open to internships, research and full-time roles
SEC.01 // About

The person behind the commits

Thesis: I am strongest where models, hardware, product constraints and deployment all collide.

My route into AI ran through hardware. I joined Technocrats Robotics as a mechanical member in my first year, and the problems that pulled me in kept moving up the stack: from the rover's frame to the vision system that decides where it can drive safely. That terrain-analysis work grew into a research paper, and into leading the rover department.

The same pattern kept repeating. A computer vision internship became four pipelines running on Jetson hardware inside client workflows. A semantic search experiment became SemantiCache. A document Q&A prototype became iORA DocQA, a shipped product. If something I build works, I want it deployed, not archived in a notebook.

Today I lead the rover department at Technocrats, sit in the founder's office for technology and AI at iora translating engineering constraints into product decisions, and have two papers under peer review. The through-lines are RAG systems, real-time vision and autonomous robotics.

Off the clock it is the gym, cricket and cooking. The foodie side leaks into the work, which is why a food-delivery intelligence system and a shared grocery cart app sit two sections down.

DegreeB.Tech CS (AI and ML), VIT Chennai, 2023 to 2027
CGPA8.28 / 10
FocusAI systems, RAG, computer vision, ML, robotics
NowFounder's Office (Technology and AI) at iora. Intern at IAMAI. Rover Lead at Technocrats Robotics
BaseIndia (Chennai / Noida)
Off-hoursGym, cricket, cooking
SignalNot a notebook-only AI profile: shipped RAG, edge CV, embedded product work and rover autonomy.
SEC.01B // Video resume

A fast pass beyond the PDF.

A quick spoken overview of how I think about AI coding workflows, product ownership and startup-paced engineering.

SEC.02 // Experience

Where I have shipped

May 2026 - Present Current

Software Testing and Product Development Intern Internet and Mobile Association of India (IAMAI)

On-site · Delhi

  • Test web applications end to end: identify bugs, check performance, document findings, and improve reliability and user experience.
  • Contribute to development of a GIS-based mapping platform: map features, level-of-detail visualization, and feature validation.
QA lens E2E flows Product lens GIS validation Outcome reliability notes that ship
Software testingSeleniumGISProduct development
Aug 2025 - Present Current

Founder's Office · Technology & AI iora

Remote · Safety-tech wearable startup

  • Sit between the founders and the technical team across embedded hardware product development, after joining as a freelance embedded systems developer (August 2025).
  • Work directly with PCB designers and firmware engineers to track progress, surface blockers early, and keep hardware and firmware workstreams coherent with product direction.
  • Review technical constraints before they become product commitments and translate engineering reality into decisions the founding team can act on.
Bridge founders to firmware Review PCB + constraints Outcome product decisions grounded in engineering
Embedded systemsPCB design reviewFirmware integrationProduct strategy
May - Jul 2025

Computer Vision Engineer, Intern Proeffico Solutions Pvt. Ltd.

On-site · Noida

  • Designed and deployed 4 production CV pipelines: speed monitoring, pothole detection, fire detection, and workplace safety analytics.
  • Ran them on NVIDIA Jetson Xavier NX at under 40 ms real-time inference latency, replacing manual monitoring with automated real-time decisions in client workflows.
4 production CV pipelines <40 ms Jetson inference Shift manual monitoring to real-time decisions
NVIDIA JetsonReal-time inferenceEdge deploymentObject detection
SEC.03 // Projects

Flagship work

Status is stated honestly: Live deployed and reachable · Access-gated deployed, but not a public demo · Sleeping demo free-tier app may need wake time · Running locally functional, not yet deployed · Demo offline code public, hosted demo down. Repo activity refreshes from GitHub.

01 // Category

Generative AI & AI Applications

iORA DocQA interface: document upload panel with file and token counters on the left, Ask and Summarize tabs with a grounded-answer area on the right
P.01 Access-gated deployment

iORA DocQA

Document question answering that picks its own retrieval strategy.

  • Problem: answers grounded in your own documents, including the numeric questions language models get wrong.
  • Built: automatic routing between direct context, Gemini + pgvector RAG, and model-written SQL executed by DuckDB on the real table; per-user corpus with cited sources and an opt-in memory layer.
  • Result: a complete shipped product with auth, persistent multi-file corpora and source-cited answers.

Owned: everything, solo build: architecture, retrieval routing, SQL path, auth, deployment.

Routing direct context / RAG / SQL Grounding cited source files Status code public, deployment access-gated
RAGGeminipgvectorDuckDBPostgreSQLRailway
P.02 Sleeping demo

SemantiCache

Semantic retrieval with a cache that understands meaning, not strings.

  • Problem: semantically similar queries keep paying full vector-search cost.
  • Built: a GMM cluster-partitioned semantic cache in front of FAISS; FastAPI service, Dockerized, LangChain and LlamaIndex compatible.
  • Result: about 40 percent fewer redundant lookups, roughly 480 ms p95 measured end to end on the 20 Newsgroups demo corpus.

Owned: everything, solo build: cache design, clustering, API, Docker packaging, demo.

-40% redundant lookups 480 ms p95 demo RAG path Stack FastAPI + FAISS + Docker
FAISSGMM clusteringSemantic cachingFastAPIDocker

Free-tier services can sleep between visits; first load can take about a minute and may require one retry.

ForgeAI

Running locally

A local-first autonomous software-engineering control plane with repo indexing, approval-gated LangGraph agents, FastAPI orchestration, SSE events, RAG search, visual review artifacts and a Next.js dashboard.

FastAPILangGraphRAGNext.js

Signal: agent systems + approval-safe engineering workflows.

RecallAI

Running locally

A local-first visual episodic memory app: teach objects from camera/upload frames, add room and nearby-object context, then search remembered objects by text, image, or both.

FastAPISQLiteReactVisual search

Signal: productized CV retrieval with local persistence.

FieldOpsEnv

OpenEnv Hackathon

A deterministic benchmark environment for autonomous field-robotics agents: structured terrain grids, hazard zones, and hard energy budgets, modeling disaster response, mine inspection and precision agriculture. Supports LLM-assisted decision-making. Packaged as a Docker Space.

RL environmentsAgentsDocker

Signal: robotics agents + constrained planning.

02 // Category

Machine Learning, Deep Learning & Computer Vision

P.03 SIH 2025 Grand Finalist

Annadata

An agricultural AI platform built around an autonomous Agribot.

  • Problem: crop disease detection that actually reaches farmers in the field.
  • Built: two-stage vision (custom YOLOv8 + ResNet-34), GPS-tagged async pipelines, and a rental platform with the Agribot's live POV feed.
  • Result: 91 percent disease classification, 97 percent soil-health prediction; SIH 2025 Grand Finale, top 5 nationally in Robotics and Drones.

Owned: the vision stack: YOLOv8 + ResNet-34 pipeline and the GPS-tagged async detection flow, within the SIH team.

91% disease classification 97% soil-health prediction Top 5 SIH Robotics & Drones
YOLOv8ResNet-34GPS pipelinesFlaskPython
CraveConnect prediction results: upgrade probability gauge, customer segment with recommended action, top drivers, SHAP global feature importance and LIME local explanation panels
P.04 Insightify 6.0 National Finalist Sleeping demo

CraveConnect

Explainable customer intelligence, built under national-competition pressure.

  • Problem: predict which food-delivery customers will upgrade, and explain every single prediction to decision-makers.
  • Built: engineered behavioral features, a tuned gradient-boosting classifier (XGBoost benchmarked), SHAP global + LIME local explanations, segmentation and a Power BI layer over a Flask API.
  • Result: 1st nationally in round one and 2nd in the ML case study at Insightify 6.0 (IIM Lucknow), on the way to the grand finale.

Owned: modeling, explainability and serving: features, classifier, SHAP/LIME layer, Flask API, within team DataQuants.

ROC-AUC 0.784 PR-AUC 0.665 Explains SHAP + LIME
Gradient boostingSHAPLIMEPower BIFlask

Apple Quality Detection

Demo offline

End-to-end ML pipeline predicting apple quality from physical and chemical features: preprocessing with robust scaling, SMOTE class balancing, a tuned Support Vector Classifier, and a Flask web interface for real-time predictions.

SVCSMOTEFlask

Signal: classical ML pipeline end to end.

Explainable Few-Shot

The engineering companion to the few-shot paper: an AI-powered few-shot medical image classification system using Prototypical Networks with a ResNet-18 encoder, trained on real hospital data, with Grad-CAM for interpretable eye disease diagnosis.

PyTorchFew-shotGrad-CAM

Signal: research implementation and interpretability.

03 // Category

Data Science & Data Engineering

IAMAI CMS

Live

A public handoff mirror for a full-stack IAMAI website and CMS: editable page blocks, structured content collections, Supabase Auth/Postgres/Storage, form APIs, media workflows, Netlify deployment and GitHub Actions smoke monitoring.

Next.jsTypeScriptSupabaseNetlify

Signal: production CMS architecture + deployment operations.

EstateOS AI

Frontend MVP

A real-estate command-center prototype with role-specific dashboards for customers, brokers, builders and management. Static fixture data today, designed to show the operating-system layer before backend/AI services are wired.

ReactTypeScriptViteRecharts

Signal: dense dashboard UX and stakeholder-specific product thinking.

Basketeer

Running locally

A collaborative shopping app for students to create shared baskets, invite others, track order thresholds in real time, and coordinate checkout across platforms like BigBasket, Blinkit and Amazon.

ReactSupabaseRealtimeEdge Functions

Signal: realtime product workflows with auth and database-backed state.

ShareKart

A quick-commerce app in the Blinkit mold with one twist explored properly: shared carts, so multiple people can fill one delivery together. TypeScript and Vite frontend on a Convex realtime backend with auth.

TypeScriptConvexVite

Signal: product UX + realtime shared state.

Smart Task Manager

Running locally

A task platform with authentication, a full REST API, PostgreSQL persistence, real-time WebSocket notifications and Pandas/NumPy analytics. Functional end to end, not yet deployed.

FlaskPostgreSQLWebSockets

Signal: full-stack API, persistence and realtime events.

SEC.04 // Research

Papers and patents

Both manuscripts are with peer-reviewed journals. Venues will be named once decisions are public.

Under peer review

Few-Shot Learning with Explainable AI

An N-way K-shot classifier for low-data medical diagnosis, reaching 93 percent accuracy on hospital-sourced eye imagery. Prototypical Networks over a ResNet-18 encoder, with LIME, SHAP and Grad-CAM attribution so every prediction is interpretable for clinical review. Shareable now: companion code. Abstract and figures follow the journal decision.

Method Prototypical Networks + ResNet-18 Explainability LIME, SHAP, Grad-CAM Caveat venue hidden until decision
Few-shot learningPrototypical NetworksGrad-CAMMedical imaging
Under peer review

Vision-Based Terrain Analysis

A real-time multi-model perception system for terrain classification in autonomous navigation, combining YOLOv8 detection with MiDaS depth estimation and deployed on embedded edge hardware. Born from rover work, hardened into research. Shareable now: the rover terrain stack it grew from. Abstract follows the journal decision.

Perception YOLOv8 + MiDaS Platform embedded edge hardware Origin rover autonomy work
YOLOv8MiDaSDepth estimationEdge deployment
Patent in process

Autonomous Patrolling Robot

A patent filing on an autonomous patrolling robot for automated security rounds. Specifics stay confidential until the filing is public.

Status in process Scope autonomous patrol rounds Disclosure limited until public filing
RoboticsAutonomous navigationPatent
Mission log // Robotics

Technocrats Robotics, VIT Chennai

Mechanical member since March 2024. Rover Lead since August 2025.

RoleRover Lead, Aug 2025 onward
SystemTerrain perception for unstructured ground
CompetitionIRC 20th worldwide · Robocon AIR 1 R1

Rover Abhimanyu

Lead for the rover department. Architected the autonomous terrain analysis stack: CV-based soil classification fused with onboard sensors for navigation across rough, unstructured ground. Missions span extreme terrain traversal, autonomous path planning, robotic arm equipment servicing, and science exploration with soil sample analysis.

International Rover Challenge

20th place worldwide at IRC 2025 out of more than 100 applying teams, and a finalist for IRC 2026. One of the most competitive collegiate rover challenges, hosted by the Space Robotics Society.

DD Robocon

All-India Rank 1 in Round 1 of DD Robocon, with a perfect 100 out of 100 in stage one, as part of the team's basketball-playing robot campaign.

SEC.07 // Skills

Tools I reach for

Skills are grouped by where they have been used, not by buzzword density.

AI, LLMs and RAG

RAG pipeline designSemantic searchVector embeddingsLangChainLlamaIndexPrompt engineeringExplainable AI (SHAP, LIME)Few-shot learning

Proven in: iORA DocQA · SemantiCache

Computer vision

YOLOv8ResNetMiDaSObject detectionDepth estimationReal-time inferenceEdge deployment (Jetson Xavier NX)

Proven in: Annadata · Proeffico pipelines · rover terrain stack

ML frameworks

PyTorchscikit-learnXGBoostHuggingFace TransformersFAISS

Proven in: both papers · CraveConnect

Backend and deployment

FastAPIFlaskDockerREST APIsMicroservicesPostgreSQL + pgvector

Proven in: DocQA · SemantiCache · CraveConnect

Data and tooling

Power BI (DAX, data modeling)PandasNumPySQLGitLinuxSelenium

Proven in: CraveConnect · IAMAI GIS work

SEC.08 // Achievements

Proof of work

TOP 5

SIH Robotics & Drones

Grand Finalist nationally after clearing college-level selection from nearly 800 teams.

AIR 1

DD Robocon R1

Perfect 100/100 in stage one with Technocrats Robotics.

#20

IRC Worldwide

International Rover Challenge 2025, among 100+ applying teams.

FINALIST

Insightify 6.0

IIM Lucknow analytics competition; 1st nationally in round one and 2nd in ML case-study round.

2025

Smart India Hackathon, Grand Finalist. Top 5 teams nationally in the Robotics and Drones category out of 500 plus, after placing in the top 3 of nearly 800 teams at college level.

2025

DD Robocon, All-India Rank 1 in Round 1. Perfect 100 out of 100 in stage one with Technocrats Robotics.

2025

International Rover Challenge, 20th worldwide. Among 100 plus applying teams. Finalist again for IRC 2026.

2025

Insightify 6.0, National Finalist. IIM Lucknow's flagship analytics competition. Team DataQuants ranked 1st nationally in round one and 2nd in the ML case-study round.

// Certifications

Network Support and Security

Cisco Networking Academy · April 2025

View certificate

NetAcad Certificate 76222e88

Cisco Networking Academy · verified issuance

View certificate

NetAcad Certificate a98ea392

Cisco Networking Academy · verified issuance

View certificate

NetAcad Certificate 382370b7

Cisco Networking Academy · verified issuance

View certificate

Getting Started with Artificial Intelligence

IBM SkillsBuild · Credly verified

View badge

Journey to Cloud: Envisioning Your Solution

IBM SkillsBuild · Credly verified

View badge
SEC.09 // Education

Education

B.Tech, Computer Science (AI and ML)

Vellore Institute of Technology, Chennai · Aug 2023 to May 2027

8.28 / 10 CGPA

Relevant coursework: Machine Learning, Deep Learning, Natural Language Processing, Data Structures and Algorithms, Probability and Statistics, Computer Vision, Database Management (SQL).

AI core: ML · DL · NLP · CV Systems core: DSA · DBMS · SQL Math core: Probability · Statistics
SEC.10 // Contact

Let us build something.

Open to internships, research collaborations and full-time roles in AI, RAG systems, computer vision and robotics. Send a note here, or email me directly.

  • Shipped RAG products end to end
  • Sub-40 ms vision on Jetson, in client workflows
  • Rover lead · IRC 20th worldwide · Robocon AIR 1 (R1)
  • SIH 2025 Grand Finalist · 2 papers under review
AI/ML Engineering Computer Vision Robotics Software RAG / Backend

Message me

RK26 // inbox relay
Mail directly