~/rajesh

rajesh@edinburgh:~$ git log --graph --all

Rajesh Kumar Kona

Software engineer. Two years building enterprise systems at Accenture, an AI product that has served 6.7 million requests, and three published papers. Now studying MSc Computer Science at the University of Edinburgh, and still building.

Explore the work GitHub

Select any commit. Arrow keys work too.

BTech, VVIT first AI course 3 × Microsoft IRJET IJCRT Rank 2 / 66 Accenture LexoraAI Promoted HEAD → Edinburgh Atlas Dissertation

commit 705c893HEAD → mainorigin/edinburghSep 2026

merge: MSc Computer Science, University of Edinburgh

Every branch merges here. Studying machine learning, ML systems, text technologies, HCI and blockchains, while continuing to build.

  • Machine Learning Systems
  • Machine Learning Practical
  • Blockchains & Distributed Ledgers
  • Text Technologies for Data Science
  • Human-Computer Interaction
Read more

$ git shortlog --summary --numbered

years of enterprise engineeringAccenture, Aug 2024 – Sep 2026
2+
production contributionsAccenture
170+
critical defects resolvedwithin SLA
20+
LexoraAI pageviewsCloudflare analytics
100K+
requests servedLexoraAI, Cloudflare analytics
6.7M+
peer-reviewed publicationsIJCRT, IRJET ×2
3
undergraduate cohort rankBTech CSE, VVIT · GPA 8.65/10
2/66

Four ways to explore

$ choose --interface

Web The full story, top to bottom. You are here. scroll Terminal A working shell. Type open lexora or neofetch. press ` Explorer Browse the portfolio as a repository, folder by folder. file tree World An island of my work. Walk it as a map, or step inside in first or third person. WASD · E · V

new Rajesh World has three views

Isometricthe whole island as a map Third personwalk it with your avatar First personthrough your own eyes

enterprise · Aug 2024 – Sep 2026

Production work, under SLA

Software Engineer (Salesforce Developer)Accenture · Bengaluru, India

  • Designed and built enterprise CRM solutions in Apex, Lightning Web Components, SOQL and Flow, cutting manual processing.
  • Built secure REST API integrations for reliable two-way data exchange between core systems and third-party enterprise services.
  • Engineered reusable components aligned to enterprise architecture standards.
  • Promoted from Associate Software Engineer to Software Engineer in roughly 21 months, ahead of the standard cycle.
production contributions
170+
critical defects resolved within SLA
20+
months to promotion, ahead of cycle
~21

The bug that never reached production

Client and project details stay confidential. This shows the shape of what happened.

Release goes out clean. The client avoids a significant business loss, and the onshore team formally recognises the catch.

Before Accenture Full-stack engineer intern at Tech Stalwart Solution (Jun–Jul 2023). Improved page-load speed by roughly 15% with React.js and co-deployed the company website on AWS.

ai · shipped · 2025 – present

LexoraAI

“Your AI companion for every document.”

lexoraai.onlineOpen-source release in preparation

A privacy-first document workspace. It reads PDFs aloud with karaoke-style highlighting and takes voice commands like “go to page three”. You can select text and ask the AI companion to explain it in plain language. There are 40+ tools, including merge, split, compress, convert, OCR, sign, redact and translate.

I built all of it myself: front end, back end, LLM integration, hosting and analytics, with Cloudflare in front. Free tools never upload your files, and server features ask first and delete everything afterwards.

pageviews
100K+
requests served
6.7M+
document tools
40+
engineer
1

Traffic figures verified in Cloudflare analytics.

Where does your file go?

Your devicebrowser tabAI companionexplains selected textIn-browser toolsmerge · split · OCRAsks firstexplicit consentServer featuredeleted afterwardsResultback on your device

Free tools run entirely in your browser. The file never leaves your device.

ai · working tree · uncommitted

Atlas

An evolving enterprise AI automation architecture.

Design and build in progress · not released

Atlas grows out of the LexoraAI work and aims at enterprise use. Agents do real work through tools, but every action is checked, risky ones wait for a person, and everything leaves an audit trail. Follow one task through the design.

  1. 01 Request A tenant submits a task in plain language.
  2. 02 LLM gateway Routes model calls per tenant, with bring-your-own enterprise LLM and API credentials.
  3. 03 Orchestrator Durable workflows built on LangGraph, so long-running tasks survive restarts and resume where they stopped.
  4. 04 Agents + MCP tools Agents act through tools exposed over the Model Context Protocol instead of ad-hoc integrations.
  5. 05 Verify & risk Each proposed action is checked and scored for risk before anything touches a real system.
  6. 06 Human approval Risky actions pause for a person to approve or reject.
  7. 07 Audit log Every decision, approval and action is recorded, so the system can be audited later.
multi-tenancypersistencebilling

$ Choose a task and press Run task.

ai · in development · 2025 – present

ownVoicz

Own your voice. Use it anywhere.

ownvoicz.comIn development · intended for open-source release
Voice ID

Voice ID sits at the centre: a voice that belongs to its owner and can be used, with permission, anywhere. Five planned pillars build on it.

Voice ID core idea

The identity primitive. Everything else asks it for permission.

Pillars are the roadmap. None is presented here as released.

ai · in development · 2026

Voice-to-Video

Say it, and see it. Then make it render fast, and prove it renders right.

GitHubRead the paper (PDF)Technical report

Speak, paste a script or drop in a document, and get a narrated video back: scenes planned from meaning, visuals chosen with a written reason, captions and credits included. Most of the engineering went into rendering, which is the one cost that grows with the length of the video.

  1. 01CaptureA recording, a pasted script or a document (PDF, DOCX, PPTX, TXT).
  2. 02UnderstandTranscript → units of meaning → scenes grouped by idea, not by sentence.
  3. 03DirectFor each scene the Visual Director picks drawn, licensed or generated visuals, and writes down why.
  4. 04GroundEvery number, date or place in a drawn visual must trace back to the source, or the scene degrades.
  5. 05ComposeA timeline timed to the narration. The voice is the only clock.
  6. 06RenderFrames drawn, encoded in 12-second segments, stitched with the narration into an MP4 plus captions.

Where the render time goes

The measured profile of a 1080p render, run through Amdahl's law. Move the sliders, or try the presets. This is a model built on the measurement, not a benchmark.

Measured
compose 77%encode 23%
Accelerated
1.00×faster overall

Pull the plug

The rule: a finished segment is a file whose name says so. Start a render, then cut the power. A simulation of the design, not a recording.

One pass
0%
12 s segments
0 / 40

Both renders are drawing the same video.

Does the GPU draw the same picture?

The rule: no channel may differ from the CPU by more than 2, and no pixel may exceed that. These frames are from a run that failed it on a GeForce GTX 1650. The run after the fix passed every scene, so it saved no images.

CPU reference (Pillow)
GPU, earlier shader
Difference: amber ≤ 2, white > 2

How the research went

  1. Profile firstOn a real 1080p timeline, composing frames took 77% of render time and x264 took 23%. Text was only 0.7% of composition. The cost was moving full-frame buffers around.
  2. Remove wasted workMost of that buffer work was redundant. Removing it took composition from 12 to 26 frames a second on the same hardware.
  3. Segment the renderTwelve-second segments, cut on exact frame indices. A finished segment is a file whose name says so, so a crash loses at most one segment, and segments run in parallel across processes.
  4. Skip hardware encoding (for now)NVENC replaces only the 23%. While composition is most of the work, Amdahl caps that at 1.3×.
  5. Move only resampling to the GPUAn OpenGL 3.3 painter does the expensive, parallel part: Lanczos-3 resampling of photographs. Text, overlays and transitions stay on the CPU reference so they match exactly.
  6. Prove it matchesNo channel may differ from the CPU by more than 2, and no pixel may exceed that. The first run on a GTX 1650 passed 15 of 23 scenes; after four fixes, 23 of 23.
  7. Measure it, then profile againOn my GTX 1650 laptop at 1080p, photographs compose 11.7× faster on the GPU (3.7 to 43.3 fps). That makes x264 the bigger cost for photo-heavy videos, so hardware encoding is worth measuring now.

Four bugs, found by counting pixels

First run on real hardware: 15 of 23 scenes. Every failing scene had something vertically asymmetric in it.

  1. Framebuffer read bottom-upA plate at rows 11–50 differed across rows 11–169: itself and its mirror.
  2. Cleared to black, not the theme background18,432 of the 20,736 differing pixels in one scene.
  3. Picture drawn at its box, not the clamped cover boxThe other 2,304 pixels, a nine-row band.
  4. Overlays alpha-composited instead of replacedOverlay scenes off by 193. Now drawn by the reference.

18,432 + 2,304 = 20,736, exactly the reported count for that scene. That is why it was a measurement and not a guess.

How much faster the GPU is

ContentCPU painterGPU painterGPU / CPUx264 encode
Photographs3.7 fps43.3 fps11.73×24.2 fps
Transitions2.1 fps16.6 fps7.71×26.8 fps
Mixed7.2 fps21.4 fps2.99×22.6 fps
Typography30.6 fps28.8 fps0.94×48.4 fps

1080p on my GTX 1650 laptop, median of three runs. Typography stays on the CPU, which is why segments are routed by content. With photographs on the GPU, x264 becomes the slower half.

Long renders

VideoRender timeSegmentsOn GPUPeak memoryMemory trendPeak VRAMLength error
120 min142.0 min600598465 MB−0.095 MB/segment505 MB0.00 s
240 min341.4 min1,2001,198471 MB−0.113 MB/segment916 MB0.00 s

GeForce GTX 1650, 1080p30. Before this, the longest video the project had rendered end to end was 150 seconds.

What is not claimed

  • GPU speed was measured on one machine (a GTX 1650 laptop, three repeats). Other GPUs will give other numbers.
  • All real-GPU evidence comes from that one GeForce GTX 1650 (plus Mesa llvmpipe, a software OpenGL).
  • The long renders used a synthetic fixture, not real footage.

Python 3.11PydanticStarletteFFmpeg / x264PillowNumPyOpenGL 3.3 (moderngl)SQLite (WAL)TypeScriptDocker

ai · open source · 2026

Salesforce AI Agent

An agent runtime for Salesforce, with the safety layer outside the model.

GitHubOpen source · MIT

An open-source agent that works on a real Salesforce org from plain language: inspect schema, query data, change records, build fields and Flows, write and test Apex, and take a release through validate, diff, approve, deploy and verify. The model proposes. A deterministic risk engine outside the model decides, and risky changes wait for the right people.

A language model is the reasoning engine, chosen per project from ten providers. Salesforce's own APIs are the execution layer: REST and SOQL, the Metadata API, the Tooling API for Apex, and Bulk API 2.0. Authorization, risk classification and approval sit outside the model, so they cannot be talked around.

tools
52
of them change something
20
automated tests
687
model providers
10
  1. Understand
  2. Inspect the org
  3. Plan
  4. Select tools
  5. Execute
  6. Observe
  7. Validate
  8. Approval, when required
  9. Execute the change
  10. Verify in Salesforce
  11. Report

Ask it to change something

Pick a request and where it lands. Each answer is what the repository's own risk engine returned for that tool, with its real declaration and the default project policy. Nothing here connects to Salesforce.

Describe the Account object

What an approval is bound to

The exact change
A SHA-256 hash of the tool name and its canonical arguments. Change one argument and the approval authorizes nothing.
The org as it was
A fingerprint of the org state the change was proposed against. If the org moved while the change waited, it is refused and proposed again.
A deadline
Approvals expire, sooner the riskier: 60 minutes for a medium-risk record change, 15 for a real deployment or a permission change, as little as 10 for critical ones.
Enough people
A quorum per category and risk tier. Permission changes and real deployments need two distinct people from eligible roles.
Not yourself
Separation of duties: the person who asked cannot approve a high-risk change when the project requires it, and never a critical one.

What it can work on

  • Schema and SOQL
  • Records
  • Metadata and fields
  • Record-triggered Flows
  • Apex: write, compile-check, test, coverage
  • Data quality and Bulk API 2.0
  • Org debugger
  • Dependency analysis
  • Reports
  • Permissions and access audit
  • Change sets: diff, deploy, verify, rollback plan
  • Jira, GitHub and Bitbucket

Of the 52 tools, 32 are LOW, 13 are MEDIUM and 7 are HIGH risk by declaration; the engine starts there and escalates on environment, object, blast radius and deletion.

What is not claimed

  • Its tests run against doubles. Integration with a real org is manual and sandbox-first, as the README says.
  • It is open-source code, not a hosted service: no payments are taken, and no SOC 2 or ISO 27001 is claimed.
  • AWS Bedrock, Google Vertex and SAML are not implemented; selecting one fails at startup instead of pretending.

Python 3.11FastAPISQLAlchemy 2 (async)AlembicPostgreSQL / SQLiteNext.js 15React 19TypeScriptSalesforce REST, Metadata, Tooling, Bulk 2.0OAuth 2.0 + PKCEOIDC · SCIM 2.0MCPDocker Compose

ai · open source · Sep – Oct 2026

Career OS

It does the searching and the typing. I do the deciding.

GitHubOpen source · MIT

A local, open-source job-hunting system. It finds jobs, internships, fellowships and scholarships worldwide, scores them, drafts the cover letter, and fills in the application in a real browser. Then it stops: I read every answer and click Submit myself. It tracks replies from Gmail and teaches me for every role I am aiming at.

Everything runs on my own laptop: a SQLite database, a local web app and a real browser window. The agents use my own AI subscription through Claude Code, Codex or Gemini, never an API key, so the whole thing costs nothing extra to run.

preset companies
112
job systems and feeds
12
learning tracks
16
applications sent for you
0
  1. 01 Find Scouts read the public job APIs of Greenhouse, Lever, Ashby, SmartRecruiters, Workable and Workday, plus four open feeds, twice a day. Research agents look for fellowships, scholarships and firms with no public API.
  2. 02 Score Rule-based scoring from 0 to 100: target role, level, country, visa-sponsorship wording, skills and deadline.
  3. 03 Prepare The AI writes a fit note, the gaps, a cover letter and a “why this company” answer, from my own facts only.
  4. 04 Fill A real browser window opens. Obvious fields are filled first, then the AI reads the whole page and drafts the rest.
  5. 05 · you You submit Nothing is ever submitted for you. Consent boxes stay unticked until you press “Tick acknowledgements”.
  6. 06 Track Gmail is read over IMAP, read-only, and replies move the pipeline. After 14 quiet days a follow-up is saved to Drafts. Nothing is sent.
  7. 07 Learn Mandatory tracks plus one per role, an AI tutor, tests with Python problems run locally, and mock interviews.

Fill a form, then stop

A replay of how the autofill marks a page. Green is something I have already answered, amber is a draft I must read, red is something only I can know. The Submit button is the site’s own, and only a person presses it.

Full name—Your answer
Email—Your answer
CV—Your answer
Why this team?—AI draft
Referee email—Needs you
Expected salary—Needs you

$ Press Run autofill.

How it is built

  • About 6,800 lines of Python (FastAPI, SQLite in WAL mode, Playwright) and a plain-JavaScript app with no build step: 72 API routes and 16 tables, served only on 127.0.0.1.
  • Thirteen agent types on a scheduler, with a cap on how many run at once and automatic recovery when the AI plan’s usage limit is hit.
  • The AI agents run the official Claude Code, Codex or Gemini command-line apps signed in with the user’s own subscription. API-key variables are removed first, so it cannot run up API bills.
  • Autofill injects a script into every frame of the page and talks back to Python. It colour-codes every field, learns from my edits, and opens a linked second form when a page says the real application is elsewhere.
  • A 9-step setup wizard, so it works for anyone, not just me: it reads your career folder and writes what it understood into an about-me file you can edit.

What it teaches

Every track has five levels: Foundations → Core → Advanced → Expert → Frontier.

Everyone

MathsProbability & StatisticsCS FoundationsDSASoftware & SystemsInterview Skills

By role

Software EngineerQuant ResearcherQuant TraderQuant DeveloperML EngineerResearch ScientistData ScientistSalesforceSlackTrading

Only free, legally open material is linked: open textbooks, OCW courses, official docs and arXiv papers. My own books stay in my own folder.

What is not claimed

  • There are no automated tests yet.
  • Set-up, sign-in and pop-ups are built for Windows first.
  • Your CV and the form text are sent to the AI app you choose, so they are only as private as that service.
  • Autofill is switched off on LinkedIn, Indeed, Naukri and Glassdoor. The optional Extra page can read your own job-site searches there, which those sites do not allow, so its defaults are slow and small.

PythonFastAPISQLite (WAL)PlaywrighthttpxJavaScriptIMAPClaude Code / Codex / Gemini CLIPowerShell

research · quant · Sep – Oct 2026

AuTrad

A trading research lab built to say no.

GitHubPaper trading only

A research platform for testing trading ideas on Indian equities and crypto under strict rules against fooling myself. Every study is pre-registered and hashed before any result, every trial is logged, failures included, and an idea must clear statistical, after-cost, after-tax, capacity and forward-test gates. So far none has, and the platform says so.

Most backtests look good because the person running them tried things until one did. AuTrad is built the other way round: decide the test first, write it down, hash it, log everything, and charge every result for the searching that went into it.

pre-registered research versions
11
experiment results logged
114
automated tests
222
proven alpha, so far
0

The gate

Strictly in order; the first failure rejects. Only an idea that reaches the end may be considered for capital, and only with my approval.

  1. 01Statisticalsignificant after Holm, right sign
  2. 02After costs≥ +2% a year net
  3. 03After taxbeats the benchmark after tax
  4. 04Double costsstill positive
  5. 05Capacitystill positive at ₹1 crore, with cost drag ≤ 1.25× that at ₹10 lakh
  6. 06Stability⅔ of subperiods positive
  7. 07Forward testpasses a frozen paper test

Every version, and what it found

Quoted from the repository’s own results, from v0.3 to v1.4. Pick a version.

Question

13 strategies, hourly to multi-year, on a sealed 2024–26 test

Finding

None qualified for paper trading.

Rules against fooling myself

Write the test before the result
Each version’s plan is written and its SHA-256 hash logged before any result exists. Later amendments are hashed and logged too.
Log every trial
An append-only ledger keeps every configuration and every failure, not just the winners.
Charge for the searching
Holm correction, Newey–West t, a deflated Sharpe ratio that counts every trial, and the probability of backtest overfitting.
Only what was known then
A company filing counts only if it was public before 15:30 IST on the decision day. Index membership is point-in-time.
Price in India
STT, stamp duty, exchange and SEBI fees, GST and capital-gains tax are modelled, so an idea has to survive what a retail trader in India would actually pay.

The one rule still running

REV1, the one rule that came closest, runs as a frozen forward test on paper from 1 October 2026: ₹10 lakh notional, a hash-chained ledger, and any change to its code voids it. It will not prove anything quickly, and the repo says so: at its information ratio a forward test can catch it breaking, not prove it.

What is not claimed

  • Paper only. The broker adapter for live trading raises an error by design, and no real money is involved.
  • The agents (company dossiers, an adversarial “10th Man” reviewer, a Main Head) are rule-based and statistical, not LLM-driven, and run in shadow.
  • REV1’s first forward decision, for the 30 September close, was computed late on 3 October, and the ledger marks it as late.
  • The whole programme so far was logged over a few days around the start of October 2026, and the forward test has barely started.

PythonpandasNumPySciPyFastAPIPostgreSQL / TimescaleDBRedisReactTypeScriptDocker

ai · ideathon prototype · Aug 2026

Voice Passport

A portable consent layer for AI voices.

GitHubDevpostLive demo

AI Passport Ideathon Participated · no award claimed. The live demo runs on free hosting, so the first load can take a minute or two.

Two voice projects, one question

Who decides how a voice gets used? Voice Passport and ownVoicz are separate projects that approach it from different sides. Pick a theme or a project.

Voice IDBoth point at the same primitive: a voice identity whose owner decides what happens to it.

Try the passport

A client-side replay of the prototype’s flow, using its seed data. The real one runs on a Node.js API.

Policy · VOICE-001 · Creator Voice #01

  • Game dialogueallowed
  • Personal projectsallowed
  • AI model trainingdenied
  • Resaledenied
  • Political advertisingdenied
  • Commercial advertisingrequires approval

Incoming requests

Authorizations

  • None yet.

Receipts

  • Every decision will leave one here.

Synthetic voices are now fast and cheap to generate, but consent has not kept up. A creator may happily license their voice for a game character yet refuse model training, advertising or resale, and when a voice moves between platforms, those choices do not travel with it.

systems · archive · BTech 2020 – 2024

Before the AI work

IoTSecurityBlockchainEnterpriseAIVoiceAI systems

BTech · published 2024

Blockchain-Based Fund Management System

Solidity smart contracts for decentralised fund allocation, with consensus-validation mechanisms and a transaction-transparency monitoring interface.

Soliditysmart contractsconsensus validation

Also published as a paper in IJCRT.

BTech · IEEE Soft-con Expo

Web3 Bookstore

A decentralised bookstore with wallet-based authentication. I designed the backend APIs and database integration.

wallet authbackend APIsdatabases

Presented at IEEE Soft-con Expo.

BTech

IoT Smart Home Security System

Facial-recognition access control wired to IoT sensors, with a backend that sends alert notifications.

IoT sensorsfacial recognitionaccess control

BTech · Innovation Fair, JNTU Kakinada

AIDER

A blockchain-integrated architecture for securing medical records, with a decentralised model for validating patient data.

blockchainhealth datavalidation

Prototype presented at the Innovation Fair, JNTU Kakinada.

research · 3 peer-reviewed publications · 1 preprint

Published work

IJCRT2024

Peer-reviewed

Blockchain Based Fund Management System

R. K. Kona

Blockchain · Smart contracts

Summary and key ideas

Secure, decentralised fund allocation using smart contracts, so that every allocation is transparent and verifiable on-chain.

  • Smart contracts as the allocation authority
  • Consensus validation before funds move
  • A transparency interface for monitoring transactions
PDF link not added yet

IRJET2023

Peer-reviewed

Blockchain and Its Applications in the Real World

R. K. Kona

Distributed systems

Summary and key ideas

An analysis of where decentralised systems hold up in practice, across finance, governance and digital trust.

  • Finance
  • Governance
  • Digital trust
PDF link not added yet

IRJET2023

Peer-reviewed

Advancements in Artistic Style Transfer: From Neural Algorithms to Real-Time Adaptation

M. B. Kona, R. K. Kona

Machine learning · Computer vision

Summary and key ideas

A survey of neural style transfer, tracing the field from the original CNN-based optimisation methods to adaptive instance normalisation for real-time use.

  • CNN-based neural style transfer
  • Adaptive instance normalisation (AdaIN)
  • Real-time adaptation
PDF link not added yet

Preprint2026

Preprint · not yet peer-reviewed

Where the Time Goes: Profiling, Segmenting and Verifying CPU and GPU Rendering in a Narrated-Video Pipeline

R. K. Kona

Systems · GPU rendering

Summary and key ideas

Where render time goes in a Pillow and x264 pipeline, and how to split the work between CPU and GPU. Composition is most of the CPU time, so hardware encoding alone is capped at 1.3×. The paper covers frame-exact segmented rendering, a GPU resampler that matches the CPU within two levels on two OpenGL implementations, and measured speed: on a GTX 1650, photographs compose 11.7× faster at 1080p, which makes the encoder the next bottleneck.

  • Per-content profile: composition 54–93% of CPU time across two machines
  • Frame-exact segments give resumable, parallel renders
  • Zero-outlier equivalence: 23/23 scenes on a GTX 1650 and on Mesa llvmpipe
  • Ablations: no clamp 23, edge-clamped taps 57, bilinear 84 levels off
  • GTX 1650 at 1080p: photographs 11.7×, transitions 7.7×, mixed 3.0× faster to compose
  • Two- and four-hour renders with flat memory and exact duration
Read paper (PDF)Paper page

main · HEAD · 2026–2027

Now: MSc Computer Science, University of Edinburgh

I came to Edinburgh with two years of production engineering behind me. These are the eight courses I'm taking this year: what each one teaches, what it expects me to be able to do at the end, and what I'm building in it. Course details come from the University's public course catalogue (DRPS).

0180 credits · all SCQF level 11

Full yearSemester 1 · this termSemester 2Summer
  • MLP Machine Learning Practical INFR11132 · Full year · 20 credits · Coursework 100%

    Lab-based deep learning: designing, implementing, training and evaluating neural networks. Semester 1 is individual coursework in Python; semester 2 is a group project in PyTorch or TensorFlow.

    What I'm building Group project: designing and training our own model.

    Learning outcomes
    • Design and implement machine learning systems
    • Read and explain technical papers
    • Design experiments with a clear methodology and evaluate the results
    • Write well-structured scholarly reports

    Builds on style transfer paper, Azure AI Engineer · Course page on DRPS ↗

  • TTDS Text Technologies for Data Science INFR11145 · Full year · 20 credits · Exam 30% · coursework 70%

    Information retrieval from the ground up: preprocessing, indexing, ranked retrieval, evaluation, web search, text classification, topic modelling and retrieval-augmented generation (RAG) with LLMs.

    What I'm building Group project: building a full search engine with a RAG layer on top.

    Learning outcomes
    • Build a search engine from scratch
    • Build feature-extraction modules for text
    • Implement retrieval evaluation scripts
    • Explain how web search engines work
    • Deliver a team project

    Builds on LexoraAI document tools · Course page on DRPS ↗

  • MLS Machine Learning Systems INFR11269 · Semester 1 · 20 credits · Coursework 100%

    The systems side of ML: data management and queries, PyTorch and GPU architecture, profiling, distributed training, deployment, inference acceleration, and safety and privacy in deployment.

    Learning outcomes
    • Explain data types and ML system architectures
    • Build and profile ML system implementations
    • Compare and evaluate systems
    • Reflect on quality and security of data and models

    Builds on LexoraAI at 6.7M+ requests, Voice-to-Video GPU rendering · Course page on DRPS ↗

  • HCI Human-Computer Interaction INFR11299 · Semester 1 · 20 credits · Coursework 100%

    Methods for understanding people, generating design ideas and assessing user experience, applied through the full HCI design cycle in a group project.

    Learning outcomes
    • Describe HCI theories and why they matter for design
    • Apply research and design methods in real settings
    • Combine insights across disciplines on complex design problems
    • Communicate research and design processes

    Builds on LexoraAI read-aloud and voice commands · Course page on DRPS ↗

  • BDL Blockchains and Distributed Ledgers INFR11144 · Semester 1 · 10 credits · Exam 70% · coursework 30%

    Distributed ledgers and the cryptography behind them: consensus, privacy, scalability, smart contracts, multiparty computation, proof of stake and space, and game-theoretic analysis.

    Learning outcomes
    • Analyse multi-party protocols and their security properties
    • Think critically about cybersecurity
    • Program smart contracts
    • Evaluate smart contract code using cryptographic primitives

    Builds on two blockchain papers, fund management system · Course page on DRPS ↗

  • IPP Informatics Project Proposal 20 INFR11291 · Semester 2 · 20 credits · Coursework 100%

    Turning a dissertation idea into a structured proposal: literature review, goals, milestones, risk and resource planning, and the legal, ethical and professional issues.

    Learning outcomes
    • Select and critically evaluate literature to justify decisions
    • Write a structured dissertation proposal
    • Plan time, resources and risk
    • Handle ethics and data-management issues

    Course page on DRPS ↗

  • ACP Applied Cloud Programming INFR11245 · Semester 2 · 10 credits · Coursework 100%

    Hands-on cloud programming, mainly in Java with some Go and Rust: containers, microservices, event processing, Kubernetes and CI/CD.

    Learning outcomes
    • Implement containerised microservices and event processing
    • Compare cloud architecture styles
    • Evaluate the major providers' offerings
    • Explain CI/CD structures

    Builds on AWS and Azure certifications, enterprise integrations · Course page on DRPS ↗

  • DISS MSc Dissertation (Informatics) INFR11077 · Summer · 60 credits · Coursework 100%

    A major piece of independent, supervised work in the final months: literature, requirements, design, implementation, experiments, evaluation and presentation.

    Learning outcomes
    • Structure and critically evaluate the knowledge around a substantial topic
    • Investigate and solve the problems that come up
    • Critically evaluate the design choices made
    • Present the work with a working demonstration

    Course page on DRPS ↗

stack trace

Every skill, traced to where I used it

Pick a skill to see the work, papers and credentials behind it.

Salesforce

Backend & web

AI / ML

Cloud & DevOps

Security & blockchain

Foundations

Select a skill.

Work

AccentureTech Stalwart internship

Product

LexoraAIAtlas (in progress)ownVoicz (in development)Voice Passport (ideathon)Salesforce AI Agent (open source)Career OS (open source)

Project

AuTrad research labFund management systemWeb3 BookstoreIoT smart home securityAIDER

Research

Paper · IJCRT 2024Paper · IRJET 2023 (blockchain)Paper · IRJET 2023 (style transfer)

Credential

Salesforce certificationsMicrosoft Azure certificationsAWS Academy + internshipCybersecurity internshipUC San Diego specializationMeta coursesAI credentials

MSc

MLP + ML SystemsBlockchains & Distributed LedgersApplied Cloud Programming

Teaching

YouTube teaching

credential vault · 20 entries

Credential vault

20 credentials

IDs and links come straight from the certificates. Salesforce certifications and superbadges link to my public Trailblazer profile, where Salesforce shows them. The Agentic AI badge is listed on my CV; its proof isn't uploaded yet.

recognition

Awards

  1. 2026

    Client recognition

    Accenture

    For catching a critical defect before it reached production.

  2. 2025

    Cheer Award

    Accenture

    Performance and cross-team collaboration.

  3. 2024

    Ranked 2 of 66

    BTech CSE cohort, VVIT

    GPA 8.65/10, the top 1% of the cohort.

  4. 2022

    Third Prize

    24-Hour Design Venture Challenge

  5. 2019

    Gold Medal

    SOF Mathematics Olympiad

community

On stage and in the room

  • PresenterIEEE Soft-con ExpoWeb3 Bookstore
  • PresenterIEEE IoT Expo
  • PresenterInnovation Fair, JNTU KakinadaAIDER
  • Organiser (volunteer)ACM events
  • ParticipantBharat Blockchain Yatra
  • ParticipantData Science Workshop, IIT Hyderabad
  • ParticipantGlobal Appathon 2026, MIT App Inventor & App Inventor Foundation (May 2026)Not among the winners · certificate of contribution for taking part and completing the participant survey
  • ParticipantAI Passport Ideathon, Devpost (Aug 2026)Voice Passport · no award claimed
  • ApplicantiQOO Hackathon 2026, Bengaluru City Battle (Aug 2026)Phase 1 · not selected for the next stage

teaching · produced in OBS Studio

Build. Learn. Explain.

If I can explain something from scratch, I understand it. I teach computer science on YouTube and document the Edinburgh MSc as it happens.

Learn DSA Daily Data structures and algorithms, from scratch. @KRK0010 Explain The Infinite Machine Computer science and technical explanations. @KRK017 Build UK Diaries Telugu The Edinburgh MSc journey, documented in Telugu. @EdinburghDiariesTelugu

interaction directory

Everything you can do here

Every interactive piece of the site, one button each. Press one and it takes you there and runs it.

Ways in

Career graph

Case studies

Evidence

Small things

$ git remote add rajesh

Let's build something.

Open to software engineering, AI/ML and research conversations, in Edinburgh or remote.

[email protected] Download CV

Built from scratch with TypeScript and Vite. No templates. Hosted on GitHub Pages.

Press ` for a shell · ⌘K for everything.