Ashwin Harish profile photo 2 on Tal
Ashwin Harish profile photo 1 on Tal
Ashwin Harish profile photo 2 on Tal
Ashwin Harish profile photo 1 on Tal

Ashwin Harish

AI/ML Engineer with an Entrepreneurial Streak | Building Production RAG/LLM Systems | 4x Hackathon Winner

Bengaluru

< 1 year of work experience

May 20264 moPresent

Bengaluru, Karnataka, India

Designed, built, and operate a production AI visual-search and B2B commerce platform for a wholesale retail client — from initial architecture through live deployment, iterative feature delivery, and infrastructure incident response. 1. Built an AI visual-search engine using DINOv3 vision transformer embeddings and Qdrant vector search; ran a controlled accuracy study (rotation, occlusion, lighting, cropping) and found that a commonly-recommended step (background removal) was actively hurting match accuracy — removed it based on measured results. 2. Designed a multi-view embedding architecture for near-duplicate product matching, indexing each product photo as an independently searchable vector rather than a single pooled embedding, cutting ranking errors to zero in controlled testing. 3. Engineered a server-enforced role-based access control system (4 roles) with field-level data redaction, verified server-side on every request, including a price-filter design that prevents range queries from inferring hidden values. 4. Built a bidirectional integration with Shopify's GraphQL Admin API: idempotent create-or-update product sync, rate-limit-aware retry with backoff, and safeguards against data loss. 5. Designed and shipped a B2B wholesale ordering system with atomic transaction handling (PostgreSQL functions) and automatic order-completion logic driven by item-level fulfilment state. 6. Load-tested the system at 100 concurrent users, validating graceful degradation with zero crashes; maintain a test suite of 100+ automated checks run before every deployment. 7. Diagnosed and resolved a production infrastructure incident (managed database auto-suspension) by building a self-healing monitoring job that detects and reverses the failure state automatically via the provider's API.
4Good.AI
Jun 20262 mosJul 2026

Bangalore Urban, Karnataka, India

1.Designing and building GitHub Enterprise Cloud architecture — org structure, automation, and access control — for a large academic institution. 2.Designed Prometheus/Grafana observability across 2 production services (7 dashboards, 65 panels) with threshold-based SLOs derived from measured traffic baselines; fixed a multiprocess metrics bug that had made 29% of API traffic invisible to monitoring 3.Built 4 production n8n automation workflows for GitHub org onboarding, integrated against the live org via API.
Greywiz
May 20262 mosJun 2026

Bangalore Urban, Karnataka, India

1.Worked on AutoPhysML, a physics-aware feature engineering pipeline that automatically maps industrial IoT sensor data to relevant physical formulae, turning raw sensor columns into ML-ready features without manual domain mapping. 2.Built the entire context-generation module: used LLM-driven prompt engineering to enrich extracted markdown with engineering variables, physical formulae, operating conditions, and domain-specific metadata, structured as JSON. 3.Built the full Retrieval-Augmented Generation (RAG) System — document chunking, metadata extraction, embedding generation, vector storage, semantic retrieval, LLM-based gating, and reranking — to recommend the correct formula for a given sensor dataset. 4.Designed a rich-context embedding strategy, bundling each formula with its surrounding domain knowledge before vectorization, which improved retrieval precision over naive chunk-based embeddings. Iteratively engineered prompts across the context-generation, gating, and reranking stages; validated the pipeline on a 200-query test set with pre-defined expected answers, achieving 195/200 (97.5%) correct formula retrieval. 5.Selected PostgreSQL with the pgvector extension for the vector store, prioritizing open-source licensing and lightweight, container-free deployment over managed alternatives
AIESEC in India
Feb 20266 mosJul 2026

Bengaluru, Karnataka, India

Responsible for supporting business development initiatives, including lead generation, outreach, and partnership development. Contribute to building relationships with organizations and stakeholders to support AIESEC programs. Collaborate with team members on communication, coordination, and execution of business development activities while developing skills in professional networking, communication, and stakeholder management.
Naviotech Solution Pvt Ltd
Dec 20253 mosFeb 2026

Bangalore Urban, Karnataka, India

• Developed a customer churn prediction model using XGBoost, improving predictive performance on tabular data • Reduced model overfitting gap from 19% to 7% through hyperparameter tuning and optimization • Achieved 76% accuracy and 76% churn recall, enabling early detection of high-risk customers • Designed and implemented a complete ML pipeline including preprocessing, training, validation, and performance evaluation GITHUB: https://github.com/ash29032006/customer-churn-predictior

Education

PES University

Bachelor of Engineering - BE, Artificial Intelligence

PES University

20244 yrs2028

Join Ashwin and 10k+ other professionals on

Tal Boss

Tal is the app where bosses meet engineers & builders directly, without middlemen

Download Tal Boss

Ashwin Harish on Tal

Headline

AI/ML Engineer with an Entrepreneurial Streak | Building Production RAG/LLM Systems | 4x Hackathon Winner

About

AI/ML undergraduate at PES University (CGPA 9.32, top 5% of cohort) with production experience across retrieval-augmented generation, LLM-based retrieval systems, and computer vision, built across three AI internships. IBM Certified RAG/LLM Architect and 4× hackathon winner, including 1st place at CIDECODE 2026 (a national hackathon judged by Criminal Investigation Department(CID) Karnataka) and 1st place among 80+ teams at the Vichaar Ideathon. Recent work: architected a GitHub Enterprise org for ~9,200 students across 2 campuses, fixed a multiprocess Prometheus/Grafana bug that had hidden 29% of API traffic from monitoring, and built a full RAG subsystem for physics-formula retrieval that hit 97.5% accuracy on a 200-query test set. Currently co-founder and technical lead on a small production visual-search project — DINOv3 embeddings, Qdrant, FastAPI, Next.js — with load-shedding under concurrency and role-based access. Comfortable across the stack: Python, FastAPI, PostgreSQL/pgvector, LangChain/LangGraph, React/Next.js, and MLOps tooling (Docker, MLflow, Prometheus, Grafana). Always happy to talk RAG architecture, production ML systems, or hackathons.

Location

Bengaluru

Work experience

Founder at AtoZ Technologies, May 2026 - Present · 4 mo, Bengaluru, Karnataka, India, Designed, built, and operate a production AI visual-search and B2B commerce platform for a wholesale retail client — from initial architecture through live deployment, iterative feature delivery, and infrastructure incident response. 1. Built an AI visual-search engine using DINOv3 vision transformer embeddings and Qdrant vector search; ran a controlled accuracy study (rotation, occlusion, lighting, cropping) and found that a commonly-recommended step (background removal) was actively hurting match accuracy — removed it based on measured results. 2. Designed a multi-view embedding architecture for near-duplicate product matching, indexing each product photo as an independently searchable vector rather than a single pooled embedding, cutting ranking errors to zero in controlled testing. 3. Engineered a server-enforced role-based access control system (4 roles) with field-level data redaction, verified server-side on every request, including a price-filter design that prevents range queries from inferring hidden values. 4. Built a bidirectional integration with Shopify's GraphQL Admin API: idempotent create-or-update product sync, rate-limit-aware retry with backoff, and safeguards against data loss. 5. Designed and shipped a B2B wholesale ordering system with atomic transaction handling (PostgreSQL functions) and automatic order-completion logic driven by item-level fulfilment state. 6. Load-tested the system at 100 concurrent users, validating graceful degradation with zero crashes; maintain a test suite of 100+ automated checks run before every deployment. 7. Diagnosed and resolved a production infrastructure incident (managed database auto-suspension) by building a self-healing monitoring job that detects and reverses the failure state automatically via the provider's API.

Artificial Intelligence Intern at 4Good.AI, Jun 2026 - Jul 2026 · 2 mos, Bangalore Urban, Karnataka, India, 1.Designing and building GitHub Enterprise Cloud architecture — org structure, automation, and access control — for a large academic institution. 2.Designed Prometheus/Grafana observability across 2 production services (7 dashboards, 65 panels) with threshold-based SLOs derived from measured traffic baselines; fixed a multiprocess metrics bug that had made 29% of API traffic invisible to monitoring 3.Built 4 production n8n automation workflows for GitHub org onboarding, integrated against the live org via API.

AI/ML Intern at Greywiz, May 2026 - Jun 2026 · 2 mos, Bangalore Urban, Karnataka, India, 1.Worked on AutoPhysML, a physics-aware feature engineering pipeline that automatically maps industrial IoT sensor data to relevant physical formulae, turning raw sensor columns into ML-ready features without manual domain mapping. 2.Built the entire context-generation module: used LLM-driven prompt engineering to enrich extracted markdown with engineering variables, physical formulae, operating conditions, and domain-specific metadata, structured as JSON. 3.Built the full Retrieval-Augmented Generation (RAG) System — document chunking, metadata extraction, embedding generation, vector storage, semantic retrieval, LLM-based gating, and reranking — to recommend the correct formula for a given sensor dataset. 4.Designed a rich-context embedding strategy, bundling each formula with its surrounding domain knowledge before vectorization, which improved retrieval precision over naive chunk-based embeddings. Iteratively engineered prompts across the context-generation, gating, and reranking stages; validated the pipeline on a 200-query test set with pre-defined expected answers, achieving 195/200 (97.5%) correct formula retrieval. 5.Selected PostgreSQL with the pgvector extension for the vector store, prioritizing open-source licensing and lightweight, container-free deployment over managed alternatives

Junior Business Development Member at AIESEC in India, Feb 2026 - Jul 2026 · 6 mos, Bengaluru, Karnataka, India, Responsible for supporting business development initiatives, including lead generation, outreach, and partnership development. Contribute to building relationships with organizations and stakeholders to support AIESEC programs. Collaborate with team members on communication, coordination, and execution of business development activities while developing skills in professional networking, communication, and stakeholder management.

Artificial Intelligence Intern at Naviotech Solution Pvt Ltd, Dec 2025 - Feb 2026 · 3 mos, Bangalore Urban, Karnataka, India, • Developed a customer churn prediction model using XGBoost, improving predictive performance on tabular data • Reduced model overfitting gap from 19% to 7% through hyperparameter tuning and optimization • Achieved 76% accuracy and 76% churn recall, enabling early detection of high-risk customers • Designed and implemented a complete ML pipeline including preprocessing, training, validation, and performance evaluation GITHUB: https://github.com/ash29032006/customer-churn-predictior

Education

Bachelor of Engineering - BE, Artificial Intelligence, PES University