21 | AI Engineer | Harnessing AI beyond its limits.
AI engineer working on LLM applications, RAG retrieval, model evaluation, and the backend services that run them — mostly Python, Go, and TypeScript, deployed on AWS and Docker.
I build the evaluation harnesses and retrieval pipelines first, because that is how a model's behaviour gets measured instead of assumed. Outside the terminal, chess, binge-watching, and travelling are my favourite ways to disappear.
Retrieval-augmented generation pipelines and language-model applications, including chunking, retrieval ranking, and the failure handling real user input requires.
Evaluation harnesses that score model behaviour, retrieval quality, and latency across fixed test scenarios, so changes are measured against a baseline instead of a hunch.
Event-driven pipelines, retries, and monitoring around AI services, containerized with Docker and deployed on AWS.
Tools and dashboards for building and evaluating AI systems — benchmarking harnesses, experiment tracking, and internal interfaces.
Lamatic.ai | Applied AI InternInvestigated agent failures in production workflows — hallucinated plans, tool-routing errors, stuck reasoning loops — and designed, A/B-tested, and deployed fixes, submitting patches to LangChain and Metaflow.
Hasprana Health Care Soln Pvt Ltd | AI EngineerAudited the confidence-scoring stage of a clinical imaging pipeline, where a geometric fit scored more reliably than the ML model's softmax output, and retrained the U-Net to raise IoU from 0.888 to 0.970 on held-out validation.
LLM evaluation and red-teaming platform. Runs dataset-driven evaluations with a pluggable metric framework and LLM-as-a-judge, and adaptive adversarial campaigns that mutate attacks between rounds. Provider-neutral across OpenAI, Anthropic, and Groq, with durable execution on Temporal, async PostgreSQL persistence, and reusable run comparisons.
Multi-stage image provenance pipeline combining face detection, ArcFace representations, reverse-image discovery, metadata extraction, and deterministic evidence fingerprinting. Integrated Google Lens discovery through SerpApi and Ethereum Sepolia anchoring, via a Solidity registry contract, to produce independently verifiable provenance records with explicit failure and blocked states.
Govt of India
Built a water-monitoring application for the Ministry of Jal Shakti that allows users to capture live photographs of water levels in flood-prone areas. These images help authorities assess local water conditions and determine appropriate response measures.
Hack2skill
Developed a concept for an agentic city-intelligence application that brings together disparate urban reports, synthesizes related events into actionable summaries, and supports multimodal citizen reporting through geo-tagged photos or videos. The challenge also explores predictive alerts and map-based visualization of emerging city conditions.
Govt of India
Explored a satellite-imagery super-resolution challenge focused on enhancing medium-resolution Sentinel-2 data for finer-scale analysis while preserving geospatial and spectral consistency. The problem emphasizes applications in agriculture, urban mapping, and disaster assessment, with validation against high-resolution references and explicit uncertainty handling.
Hack2skill
Built Amanat AI, an AI-oriented document-verification platform designed around document uploads, document processing, verification workflows, and status tracking. The project is intended to streamline document-verification processes through a modern web interface and extensible AI/OCR integrations.
Ernest & Young (EY)
Designed an agentic retail-sales solution concept centered on a conversational Sales Agent coordinating specialized agents for product recommendations, inventory, promotions, payments, fulfillment, and post-purchase support across online and physical channels.
Amazon India
Participated in Amazon's 2026 machine-learning competition, a multi-source business entity resolution task: given business records from three independent sources with noisy, inconsistent names and addresses, match each secondary record to the correct deduplicated reference entity. The dataset spans tens of millions of records and results are scored with a precision-weighted macro-averaged F₀.₅ metric (per the problem statement). The reported score for this submission was 0.9482.
WeMakeDevs
Built TRON, an AI-powered organizational intelligence platform using Go and FalkorDB to connect organizational data, retrieve relevant evidence, and support grounded decision-making through graph-based reasoning.
A Comparative Study of Phi-3-mini, Mistral-7B, and Gemma-2 on SQuAD v2
Benchmarks small language models for domain-specific question answering. Compares Phi-3-mini, Mistral-7B, and Gemma-2 on SQuAD v2 using Exact Match and F1 evaluation, while examining inference time and the generative–extractive mismatch in instruction-tuned models.
Compared three small language models under identical zero-shot conditions on SQuAD v2, evaluating 200 examples with Exact Match, F1, and inference time. The results show an accuracy–latency trade-off with no model dominating, and a generative–extractive mismatch that penalises correct full-sentence answers under extractive scoring.
View Research PaperValidation: 219 terminal-related tests passed; ESLint clean.
Have something interesting in mind? Let's talk.