Available for work

Hello

I am Mohan
AI/ML Engineer

AI/ML engineer based in Surrey, UK. I don’t stop at the notebook — I own the whole path, from the first experiment to the RAG pipeline running in production under 200 ms.

Download CV Let’s connect
MSc AI · University of Surrey/ Distinction dissertation/ 12 shipped projects

// About me

The person behind the pipelines

Mohan Sree Vijayakumar
Mohan Sree Vijayakumar AI Engineer & Automation Consultant
By the numbers
— 

--:-- in Surrey · —

Currently

Building an in-house agentic GenAI sales assistant — LLM APIs wired into market-research tooling, with progress reported straight to the CEO.

Workforce Training & Development · London
0Projects shipped
0Engineering roles
0Inference latency
0Best model accuracy
Daily drivers
PythonPyTorchFastAPIAgentic workflows FAISSTransformersDockerAWSReact

// Featured works

The ones that taught me the most

01 4 months · 2025

Explainable fraud detection over unstructured text

Guardian AI needed fraud scenarios classified from free-text reports without the model inventing anything. I built a RAG pipeline over FAISS with structured output validation, then served it through FastAPI microservices for high-throughput, low-latency inference.

Guardian AIAI Developer
RAGFAISSPrompt engineeringFastAPI
View repo ↗
02 6 months · 2026

Alzheimer’s biomarkers from 73GB of single-cell data

MSc dissertation, graded Distinction. An end-to-end framework over scRNA-seq: 97.96% Random Forest accuracy, a spatial GNN for cell-to-cell interaction, and SHAP analysis that surfaced APOE and TREM2 as drivers. Fourteen age-matched files, 73GB, processed in under an hour.

University of SurreyResearcher
PyTorchGraph Neural NetsCVAESHAP
Read the paper ↗
03 Hackathon · 2025

An agent that sources candidates 10× faster than a human

Recruitment automation built on async multi-platform sourcing — LinkedIn, GitHub, Stack Overflow — scored against a six-dimensional match rubric. 94% matching accuracy, 100+ concurrent searches. Runner-up at the Synapse AI hackathon.

Synapse AI HackathonSolo build
AI agentsGeminiOpenAIAsync Python
View repo ↗
04 Final year · 2024

A virtual fitting room built on GANs

Hyper-realistic eyewear simulation for fashion e-commerce, with AR integration across mobile and web. Better product visualisation cut expected return rates by 35%. The work was published.

Anna UniversityPublished research
GANsTensorFlowOpenCVAR
Read the paper ↗
Eight more on GitHub ↗

// Open source

Fixing gradients in other people’s repositories

Every pull request and issue I have filed to repositories I don’t own. The bulk is numerical-correctness work in differentiable computer vision — defects that pass every test and every eyeball: a forward value that is exactly right while the backward pass returns nan, an eps floor that underflows in half precision, a torch.where that differentiates the branch it discards.

0Pull requests
0Repositories
0Merged
0Issues filed
The ledger

18 entries, newest first

  1. kornia #4184 Merged 04 Sep 2026

    Restored Boxes3D.to_tensor gradients after a four-year false positive

    The guard had raised on any input requiring grad since 2021, blocking all differentiable use of Boxes3D. The original report was a gradcheck artefact — a genuine kink where numerical and analytical gradients both hold and disagree.

    false-positive gradcheck+106 −48 · 3 filesView PR ↗
  2. kornia #4158 Merged 03 Sep 2026

    Guarded distance_transform against kernel underflow

    When the farthest taps of exp(-dist/h) underflow to exactly zero, the convolution drops them and the function returns a wrong — or silently all-zero — distance field. Reachable from the documented defaults.

    silent underflow+126 −34 · 3 filesView PR ↗
  3. kornia #4156 Merged 02 Sep 2026

    Largest-remainder apportionment for MultiResolutionDetector quotas

    Feature quotas were split proportionally across pyramid levels then truncated independently, so a small num_features collapsed onto a single scale. Largest-remainder apportionment spends the whole budget.

    quota truncation+48 −11 · 3 filesView PR ↗
  4. kornia #4124 Merged 31 Aug 2026

    Documented match_fginn’s unused lafs1 argument

    A required positional parameter the body never reads, kept for API symmetry with match_adalam. Closed the remaining half of a docstring report.

    API documentation+3 −1 · 1 fileView PR ↗
  5. kornia #4246 Approved 05 Sep 2026

    Corrected quaternion_to_axis_angle’s identity gradient

    At the negative unit identity autograd returned +2 where central differences give -2. Fixed to the 2/w limit with the atan2 backward shielded across the whole zero-vector branch. Maintainer-approved on the exact head, CI green.

    wrong-sign gradient+136 −3 · 3 filesView PR ↗
  6. kornia #4228 In review 04 Sep 2026

    Finite gradients at the rotation singularities of the quaternion conversions

    Both differentiate an inverse trig function at its domain boundary and multiply the infinity by an exactly-zero vector part, so every gradient through them was nan. What changed in torch 2.14 is clamp’s backward, which only masks the defect.

    0 × inf → nan+126 −4 · 3 filesView PR ↗
  7. kornia #4405 In review 10 Sep 2026

    Finite gradients at the identity for every Lie group exp and log

    So3, Se3 and Se2 all returned nan gradients at the identity on every dtype — the standard initialisation for pose optimisation. One mechanism throughout: torch.where differentiates the branch it discards. Forwards verified byte-identical against main.

    where-branch poisoning+121 −20 · 7 filesView PR ↗
  8. huggingface/peft #3706 Pooled 09 Sep 2026

    RoAd corrupts the base weight before safe_merge can reject

    merge(safe_merge=True) commits the merged weight and only then validates the bias — the exact outcome safe_merge exists to prevent. Filed with a reproduction, then fixed. The maintainer pooled seven concurrent safe_merge PRs into one fix; a queue decision, not a rejection.

    write-before-validate+95 −5 · 4 filesView PR ↗
  9. huggingface/lighteval #1371 In review 30 Aug 2026

    Honour concurrent_requests for the OpenAI and TGI judge backends

    The litellm backend read the option; the path shared by openai and tgi hard-coded ThreadPoolExecutor(10) and never looked at backend_options at all.

    +4 −2 · 1 fileView PR ↗
  10. huggingface/lighteval #1352 In review 22 Aug 2026

    Fixed broken code examples in the custom-model docs

    Re-verifying a stale report found the reported breaks already fixed, but two new ones in the same file: an unquoted task string that is a SyntaxError, and a CLI example with an unclosed quote.

    +2 −2 · 1 fileView PR ↗
  11. torchmetrics #3474 In review 20 Aug 2026

    Documented BinaryAUROC’s per-batch logit detection

    update() decides per batch whether predictions look like logits. With small batches, some are treated as probabilities and others as logits in the same run. Documented rather than changing logic other metrics depend on.

    +7 −0 · 1 fileView PR ↗
  12. langchain #38817 Policy-closed 12 Jul 2026

    Pydantic schema support in HuggingFace with_structured_output

    The method raised NotImplementedError for Pydantic schemas under its own default method, contradicting its docstring. Auto-closed by the repository bot, which requires external PRs to link a pre-approved issue — never reviewed on merit.

    +36 −7 · 2 filesView PR ↗
  13. agent-inbox #185 In review 22 Aug 2026

    Superhuman-style hotkeys for thread navigation

    e closes the current thread, arrows move between them — suppressed inside inputs, textareas and contenteditable, and while a modifier is held. Scoped to the two shortcuts the issue specified.

    +52 −0 · 1 fileView PR ↗
  14. griptape #2290 Changes requested 27 Aug 2026

    Enabled Ruff B009 across the codebase

    Removed B009 from the ignore list and reviewed all 15 resulting getattr(x, "const") rewrites across drivers, memory and mixins.

    +21 −16 · 8 filesView PR ↗
  15. outlines #2025 In review 27 Aug 2026

    Noted the local-runtime exception bypass in llamacpp

    Every other local-runtime model file carries a docstring note explaining that it intentionally bypasses provider-error normalisation. llamacpp.py was missed when the hierarchy landed.

    +5 −1 · 1 fileView PR ↗
  16. openGym #23 In review 24 Aug 2026

    Live pose overlay for form checking

    A Form check button opens the front camera and overlays a live pose skeleton using on-device MoveNet via tfjs — no video frame leaves the browser. Scoped as the MVP the issue asked for.

    +784 −40 · 8 filesView PR ↗
  17. openGym #24 In review 24 Aug 2026

    Stopped tracking data/, which held a live session secret

    The directory was committed with a real session secret, a user database with passkey credentials, and VAPID keys. The server only generates a fresh secret when the file is absent, so every deployment built from the repo shared one attacker-readable signing key.

    committed secret+1 −25 · 4 filesView PR ↗
  18. openGym #25 In review 24 Aug 2026

    Fixed the web service’s stale Dockerfile path

    docker-compose.yml pointed at a path that does not exist, so docker compose up --build failed immediately for anyone building from source.

    +1 −1 · 1 fileView PR ↗

// Experience

Where I’ve worked

@mohansree14 loading contributions…
LessMore

  1. Workforce Training & Development

    AI Engineer & Automation Consultant

    Jun 2026 — Present London, United Kingdom Full-time · 4 mos

    Building the company’s in-house agentic GenAI sales assistant while delivering its Level 4 AI apprenticeship, with progress reported straight to the CEO.

    • Build an in-house agentic GenAI sales assistant that wires LLM APIs into market-research tooling, reporting progress directly to the CEO.
    • Deliver the Level 4 AI & Automation Practitioner apprenticeship as trainer — ML principles, automation tooling, data handling and responsible AI governance, to cohorts across industries.
    • Contribute training content to the Level 5 AI Leadership programme, translating AI capability and limitation for business owners, CEOs and directors.
    • Mentor the team’s intern through implementation and debugging while staying hands-on across the build; own project scope and documentation.
    • Guide learners end to end — from analysing a business process to deploying the AI workflow that replaces it.
    • Programme delivered by an Ofsted Good and ISO 9001 certified training provider.
    Python LLM APIs Agentic AI n8n Zapier Make RAG AI governance

  2. Guardian AI

    AI Developer

    Oct 2025 — Jan 2026 Remote Contract · 4 mos

    RAG pipelines over FAISS that cut hallucination on enterprise document classification.

    • Architected agentic workflows and RAG pipelines for unstructured text analysis, tuning FAISS vector retrieval for relevance.
    • Implemented fine-tuned LLM logic for document classification, holding high-fidelity, compliant outputs inside an Agile innovation lifecycle for enterprise clients.
    • Built scalable model-serving on FastAPI microservices, targeting low-latency inference under high throughput.
    • Engineered a Python fraud-detection interface classifying fraud scenarios from free text, using prompt engineering and structured output validation.
    Python FAISS RAG FastAPI Docker Fine-tuning Agent tooling

  3. Soft4Hire

    Machine Learning Engineer — Intern

    Mar 2024 — Nov 2024 Chennai, India Internship · 9 mos

    Hybrid recommender that lifted engagement 40% and cut the cold-start problem by 60%.

    • Built a hybrid recommendation engine combining collaborative filtering (LightFM) with content-based filtering (TF-IDF + BERT).
    • Deployed real-time inference on AWS Lambda with S3, holding <200 ms latency for personalised recommendations.
    • Implemented an A/B testing framework — 40% increase in user engagement metrics.
    • Containerised ML services with Docker for consistent development-to-production workflows.
    • Content-based fallback for new users drove 25% of all platform content consumption.
    Python LightFM BERT AWS Lambda AWS S3 Docker scikit-learn

  4. DLK Technology

    Machine Learning Engineer — Intern

    Jun 2023 — Aug 2023 Chennai, India Internship · 3 mos

    Forecasting and anomaly detection running live on a B2B analytics platform.

    • Developed sales forecasting models using time-series analysis for a B2B analytics platform.
    • Implemented anomaly detection algorithms to flag unusual sales patterns.
    • Deployed predictive models on AWS infrastructure for real-time business intelligence.
    • Built data pipelines for processing and analysing large-scale B2B sales data.
    Python pandas scikit-learn Time series Anomaly detection AWS

  5. Acmegrade

    AI Research Intern

    Mar 2023 — May 2023 Bangalore, India Research · 3 mos

    100,000+ movie reviews processed into a hybrid recommendation engine.

    • Researched and implemented recommendation algorithms for a movie platform using NLP-based content analysis.
    • Developed content-based filtering systems over textual reviews and movie descriptions.
    • Improved recommendation accuracy through a hybrid collaborative and content-based approach.
    • Processed and analysed 100,000+ reviews with natural language processing techniques.
    Python NLP scikit-learn pandas MongoDB
Toolkit

53 tools I reach for

Python PyTorch TensorFlow Keras Hugging Face scikit-learn OpenCV LangChain Gemini DeepSeek MLflow OpenAI FAISS RAG Agentic workflows SHAP YOLO FastAPI Flask Node.js AdonisJS Docker CI/CD AWS Lambda AWS S3 AWS EC2 Azure AI Microservices PostgreSQL MongoDB Redis Prisma pandas Polars NumPy pgvector BullMQ React TypeScript JavaScript Three.js WebGL Streamlit Plotly Power BI Git GitHub Jira Jupyter Colab Unity R OpenXR

// Education

Where I learned it

  1. University of Surrey

    MSc Artificial Intelligence

    2025 — Feb 2026 Guildford, United Kingdom Education · Completed

    Dissertation graded Distinction: hybrid deep learning and graph neural networks for Alzheimer’s prediction on scRNA-seq data.

    • Modules: Deep Learning, Natural Language Processing, Computer Vision, Machine Learning.
    • Built a spatial GNN for cell-to-cell interaction and a CVAE over single-cell genomics; SHAP analysis surfaced APOE and TREM2 as biomarkers.
    • Processed 14 age-matched files, 73GB, in under an hour; 97.96% Random Forest accuracy.
    • Separate Distinction (90%) for AR/VR coursework — a Meta Quest 2 game built in Unity with OpenXR and real-time mesh slicing.
    PyTorch Graph Neural Nets CVAE SHAP scikit-learn Unity OpenXR

  2. Anna University

    B.Tech Artificial Intelligence & Data Science

    2021 — 2024 Chennai, India Education · First Class

    Final-year project published: an AI-driven virtual fitting room built on GANs.

    • Modules: Machine Learning, Deep Learning, Numerical Methods, Data Analytics, Statistical Analysis.
    • Built hyper-realistic eyewear simulation with AR integration for fashion e-commerce, cutting expected return rates by 35%.
    • Certified in Python for Data Science, Fundamentals of Data Science (Phases 1 & 2) and R for Statistical Computing.
    Python TensorFlow Keras OpenCV GANs AR

// Services I provide

I can help you with these things

  1. 1.RAG & agentic systemsRetrieval that actually retrieves, agents that finish the job, evals so you know when they don’t.
  2. 2.Production ML engineeringModel to endpoint: FastAPI serving, containers, AWS deployment, latency budgets you can hold.
  3. 3.Computer vision & NLPClassification, detection, token tagging, video understanding — benchmarked, not vibes.
  4. 4.Workflow automationMake, Zapier and n8n stitched to LLM APIs, so the boring half of the business runs itself.
  5. 5.AI training & consultingLevel 4 and Level 5 programmes delivered. I can explain this to your engineers or to your board.

// Receipts

Things that are on the record

MSc Artificial Intelligence — University of Surrey Dissertation: Distinction Distinction (90%) — AR/VR coursework Runner-up — Synapse AI Hackathon Published: GAN-driven virtual fitting room B.Tech AI & Data Science — First Class Python for Data Science R Programming for Statistical Computing Fundamentals of Data Science — Phase 1 & 2

// How it works

No agency in the middle — just me, in on your product

Available

Hire me

I take a small number of projects at a time, so yours gets full attention rather than a slice of it.

Mohan Sree VijayakumarSurrey, United Kingdom
5Roles
12Projects
3Countries
Start a conversation
Engagement

Project or retainer

Scoped per project— tell me the problem first

  • One workstream at a time
  • Weekly shipping, not monthly decks
  • Code and documentation delivered together
  • Direct line to me, no account manager
  • Pause or stop whenever it stops being useful
See how to book

// Let’s talk

Pick the shape of the conversation

Choose the one that fits and it carries straight through to the message below — I reply with times the same day.

Opens your mail client — no server, no tracking. Prefer email? mohansreevijayakumare@gmail.com