Mohith D K

AI/ML Engineering Intern — A.P. Moller Maersk

On-site · Jan 2026 – present

• Built and maintained an AI-powered document extraction pipeline (GPT-4o vision + Azure Document Intelligence OCR) that auto-extracts payment data from proof-of-payment documents to release cargo shipments, processing ~$800M USD in transaction volume across 10+ countries (Philippines, Japan, South Korea, China, Hong Kong, Macau, and others).
• Designed a two-pass LLM extraction architecture with confidence-scoring and automated accept/human-review routing, achieving 90%+ extraction accuracy while keeping manual review load low at scale.
• Diagnosed and fixed production data-integrity bugs in financial field extraction (incorrect reference/date/account-number values) through root-cause analysis, adding deterministic validation safeguards against LLM hallucination; extended the platform to new markets end-to-end (country-specific logic, new fields, schema/database changes) without disrupting existing country pipelines.
• Built Grafana dashboards for observability and Hedwig-based governance with proactive alerting for the product.

Stack: GPT-4o vision · Azure Document Intelligence · LLM pipelines · Grafana · Hedwig

AI Intern — SpectoV

Remote · Jun 2025 – Jul 2025

• Created an intent engine to identify customer intents and built conversational AI modules for customer feedback analysis, enabling intelligent ticket handling and smoother customer-support workflows.
• Fine-tuned a BERT-based NLP pipeline (F1: 0.91) to automate sentiment and priority detection in customer tickets, streamlining support operations and enhancing scalability for high-volume service environments.

Stack: BERT · NLP · Conversational AI · Intent detection

Research Intern — RAPID Research Lab

On-site · May 2025 – Jun 2025

• Conducted research on "Temporal Belief Dynamics in BERT", employing a graph-theoretic approach to analyze interpretable representation drift for a publishable AI/ML project.
• Fine-tuned BERT over 10 epochs and built graph-theoretic network analyses and 2D embedding visualizations to trace representation drift, contributing to a paper accepted at ICCIS 2025.

See ~/Projects/Temporal Belief Dynamics in BERT.

Provenance

Ensemble AI-content detection suite · Oct 2026 · provenance.mohithog7.workers.dev

What it does

• Built an ensemble AI-content detector (GPT-2 perplexity, conformally calibrated classifier, C2PA, watermarking) that surfaces per-sentence attribution and method disagreement instead of a single confidence score.

Attack Lab

• Added an Attack Lab that stress-tests each detector with live structural and T5 paraphrase attacks, plus a semantic retrieval ledger (fastembed, SQLite).
• FastAPI + Next.js stack with batch mode and Docker deploy config.

Stack: GPT-2 · conformal prediction · C2PA · watermarking · T5 · fastembed · SQLite · FastAPI · Next.js · Docker

Il Colosseo

Real-time multi-agent AI debate arena · Aug 2026 · ilcolosseo.vercel.app

What it does

• Built a real-time multi-agent AI debate platform with 3 agents (two Gemini debaters arguing opposing sides plus an AI judge) over three rounds, grounded in live web research (Tavily) and streamed via FastAPI and SSE.

Quality layer

• Engineered a quality layer (topic classifier, source-integrity gate, speech evaluator) and a deep-research pipeline for grounded, citable arguments.
• Added multilingual debates in 10 Indian languages with voice output; deployed the frontend on Vercel.

Stack: Gemini · Tavily · FastAPI · SSE · Vercel

DawaiSathi

Multilingual AI medicine explainer · Jul 2026 · dawaisathi.vercel.app

What it does

• Launched an MVP of a consumer-facing pharma service (10+ users): snap a strip or type a medicine name to get plain-language usage, side effects and warnings in 11 Indian languages via Gemini, streamed over SSE.

Product

• Developed the full-stack product with Next.js 14, Supabase (PostgreSQL, Google OAuth) and Razorpay payments with tiered pricing; added follow-up medicine chat, explanation history, and Vercel deployment.

Stack: Gemini · Next.js 14 · Supabase · PostgreSQL · Google OAuth · Razorpay · SSE · Vercel

Temporal Belief Dynamics in BERT

Research project · Jun 2025 – Jul 2025

What it is

A study of how BERT's internal "beliefs" change while it is being fine-tuned.

• Mapped the evolution of BERT's internal "beliefs" across 10 fine-tuning epochs, using graph-theoretic networks and 2D embeddings to uncover and visualize hidden representation drift with research-grade clarity.
• Findings accepted for publication at the 7th International Conference on Communication and Intelligent Systems 2025.

Publication

M.D.K. et al., "Temporal Belief Dynamics in BERT: A Graph-Theoretic Approach to Interpretable Representation Drift," published at the 7th International Conference on Communication and Intelligent Systems (ICCIS 2025), BITS Pilani, Goa, Springer LNCS.

Where it came from

This work started during my research internship at RAPID Research Lab (May – Jun 2025).

Stack: BERT · graph-theoretic networks · 2D embeddings