Building Applied AI
with measurable impact.
Tanish Jindal - AI graduate from the University of Manchester, delivering applied machine learning across NLP, search, predictive analytics and computer vision. Human-centred, interpretable and business-aware.
About
Engineer, researcher, builder.
I'm an AI graduate from the University of Manchester delivering applied machine learning solutions across NLP, search, predictive analytics and computer vision. My work combines Python, SQL, data engineering and model evaluation with strong academic performance and hands-on product thinking.
I've built AI prototypes for candidate support in medical education, real-estate analytics, enquiry classification, feedback analysis and real-time road anomaly detection. I bring an interest in finance AI, an interpretability focus and business awareness to decision systems.
I'm ready to build scalable, human-centred AI that creates measurable impact.
Engineering interests
Financial AI · intelligent risk modelling · algorithmic trading research · robotics, perception systems & autonomous navigation · responsible AI & explainability · open-source ML · human-centred accessibility tech.
Outside the terminal
Guitar, music and cinema. Anime and sci-fi storytelling for world-building, pattern recognition, and thinking about how technology reshapes the people who use it.
Flagship Work
Intelligent systems, shipped.
Virtual White Stick
GitHub RepoMSc dissertation delivering real-time road anomaly detection on a curated 16,476-image RDD2022 subset with 16,400 annotated defect instances across 5 classes. A modified YOLOv8m adds a 4th CLAHE + Sobel edge channel and lightweight edge attention, with multi-pass IoU consolidation and an assistive text-to-speech prototype.
Fraud Detection API
GitHub RepoGPU-aware fraud detection system on 284,807 credit-card transactions with fraud at 0.1727%. The production-style FastAPI exposes four scoring endpoints and returns probability, decision, risk label, threshold and model version, supported by SHAP, MLflow, a PyTorch autoencoder, Docker and 5/5 passing tests.
Explainable & Fair Credit Risk Scoring
GitHub RepoMaster's-level explainable credit risk system predicting loan default from 10+ borrower / loan / credit-history features. Compared 4 model families across 7 metrics, built two pipelines (with / without loan_grade) to assess feature leakage, and shipped the more defensible no-loan-grade model with threshold optimisation, probability calibration, SHAP global/local explanations, fairness analysis across 5 segments, and a Streamlit dashboard.
GARCH Volatility Forecasting on Stock Returns
GitHub RepoEnd-to-end Python pipeline fetching, processing and modelling 6 years of AAPL daily returns, comparing ARIMA (return prediction) against GARCH(1,1) (volatility prediction) to test volatility clustering and mean reversion. ARIMA improved only ~1% over a naive baseline (AR/MA coefficients not significant, p > 0.69), consistent with market efficiency; GARCH(1,1) showed high persistence (α + β ≈ 0.96) and pinpointed peak volatility on 2020-03-17, the height of the COVID crash, using no external information. Identified a flawed initial evaluation (single-day squared returns as a noisy realized-volatility proxy) and corrected it with a smoothed proxy and QLIKE loss, revealing GARCH beats a naive rolling-volatility baseline. Grounded in Engle (1982), Bollerslev (1986) and Andersen & Bollerslev (1998), with the full methodology and correction documented in a public writeup.
CIFAR-10 - Classical ML vs. CNN
End-to-end object recognition pipeline on CIFAR-10 (60,000 images, 10 classes). Classical baseline (HOG/SIFT + Bag-of-Visual-Words + Linear/RBF SVM) tuned via grid search reached 78% validation accuracy; a CNN with 4 conv, 2 max-pool and 3 dropout layers improved to 91% - a 13-point lift with a 10-point reduction in the overfitting gap.
Experience
Where the work has shipped.
Student Consultant & Team Lead - AI Strategy
-Planily- Led a client-facing consulting project for Planily, translating school emails, PDFs and WhatsApp messages into practical AI recommendations for tasks, reminders and calendar actions.
- Designed a five-layer AI routing strategy spanning rules, retrieval, SLMs, LLM escalation and human review, prioritising each use case by business value, cost and risk.
- Evaluated Phi-4-mini, Gemma and Mistral Ministral, and compared RAG with fine-tuning for stable structured-extraction workloads.
- Modelled AI usage and cost across 50, 100, 500 and 1,000 active families using an assumption of 40 messages per family per week.
- Delivered a risk-aware three-year roadmap covering API adoption, smaller-model introduction and self-hosting thresholds, with stronger controls for payments, medical, safeguarding and child-location data.
AI Intern
-Aspire2 PLAB- Built an AI-enabled candidate support prototype with Python, LangChain, RAG, FAISS, OpenAI API and Claude Code with MCP servers, indexing 100-150 PLAB resources and FAQs.
- Developed a scikit-learn text classifier using TF-IDF / SBERT features, correctly categorising approximately 85 out of 100 validation enquiries across 5+ support classes.
- Analysed 120-150 feedback forms and enquiry records to identify recurring learning gaps and strengthen topic planning.
- Created a Streamlit / Power BI and SQL analytics dashboard, iterated with staff feedback, saving an estimated 6-10 hours per month.
AI Intern
-Heritage Insights Limited- Built a real-estate analytics pipeline (Python, Pandas, SQL) integrating 1,000-1,200 property records from Rightmove, internal and public datasets.
- Random Forest property price estimation across 2 regional markets with R² of 0.72-0.78.
- Engineered 10-12 commercial real-estate features for valuation and investment decisions.
- Automated Python, SQL, scikit-learn and Power BI / Plotly workflows, saving 8-12 analyst hours per month and supporting non-technical investment discussions.
AI/ML Consultant (Freelance)
-Family Business Portfolio- Delivered 12 AI/ML use cases across three family-owned businesses in trading, jewellery retail and rice manufacturing, working alone from data cleaning through to dashboards and automation, using Python, XGBoost, OCR, YOLOv8 and CLIP-style embeddings.
- Cut weekly stock review at the trading company from 6 hours to 90 minutes and invoice entry from 5 minutes to under 90 seconds, using demand forecasting and OCR invoice extraction that also caught 14 duplicate or inconsistent invoices.
- Surfaced ₹22 lakh of credit exposure by sorting 85 customer accounts into low, medium and high risk, and flagged 5 high-risk suppliers with 0-100 risk scores, giving the business evidence to renegotiate terms with 3 vendors.
- Cut visual product search for a jewellery retailer from 15 minutes to under 2 minutes with a CLIP-style tool over 500 product images that returns the top 5 similar items, and flagged 19 underpriced or inconsistent products across 120 test items with a pricing model.
- Built a YOLOv8 rice-quality inspection tool in Streamlit that correctly classified 151 of 180 validation images and cut sample inspection from 10 minutes to under 3 minutes per batch, then flagged 4 high-risk machines and 11 maintenance issues ahead of scheduled inspection with a risk-scoring dashboard covering 12 machines.
Education
Academic foundation.
MSc Artificial Intelligence
BEng Computer Science & Engineering
Skills
Core competencies.
Programming & Software
Machine Learning & AI
Deep Learning & CV
NLP & Text Mining
Data Science & Analytics
Databases & Data Engineering
Soft Skills
Certifications
Continuous learning.
01Artificial Intelligence & Machine Learning
8
Artificial Intelligence & Machine Learning
802Cloud & DevOps
5
Cloud & DevOps
503Finance
3
Finance
304Web & Full-Stack Development
2
Web & Full-Stack Development
205Foundational & Other
2
Foundational & Other
2Contact
Let's build something measurable.
Based in Manchester. I respond fastest to email - happy to chat about full-time roles, collaborations, or applied research in ML, computer vision, and financial AI.