I embed with product and account teams and ship production AI: LLM pipelines, retrieval systems, agent workflows and the full-stack platforms they run inside.
~/dinesh $
Not slideware, running systems. I take AI from a rough idea to something your team logs into every day.
I sit with the team that will actually use the software, work against the real workflow instead of a written spec, and own the result after launch. Scoping, building, deploying and correcting course happen in the same week rather than across quarters.
Retrieval-augmented generation grounded in your own data, multi-agent orchestration with LangGraph and CrewAI, structured extraction with JSON repair, LoRA and DreamBooth fine-tuning, and multi-provider routing across Gemini, Claude and GPT models with fallback chains and per-run cost tracking.
Multi-tenant products end to end: React 19 and Vite on the front, FastAPI and PostgreSQL behind it, OAuth with revocable JWT sessions, a live permission matrix, Celery and Redis for background work, and Docker Compose behind nginx with GitHub Actions doing the shipping.
I'm an engineer who likes being where the problem lives. I started in research at Samsung R&D (SRI-B), where I built a Video Q&A chatbot that read transcripts at 96% accuracy, then led a platform-agnostic Stable Diffusion pipeline that won the Outstanding and Excellence Awards.
Today at Trilliant Digital I work as a Forward Deployed Engineer, embedded with client account teams. I built Opaeron from the first commit: 226,000 lines, 711 API endpoints, 77 Postgres tables and about 20 AI agents running in production. Published at Springer Nature (SmartCom 2025) along the way.
Python
TypeScript
React 19
FastAPI
PostgreSQL
Redis
Celery
Docker
Nginx
Vite
LangChain
PyTorch
TensorFlow
HuggingFace
OpenAI
Gemini
Anthropic
n8n
GitHub Actions
GitFrom research benches at Samsung R&D to forward deployed client work at Trilliant Digital.
The kind of thing forward deployed actually means. I designed, built and operate this in production for client pods.
The internal operations platform for a digital marketing agency. Paste a client meeting transcript and an LLM pipeline pulls out action items with owners and dates, writes the meeting summary, and pushes everything into task boards, workload analytics, billing scorecards and a client-facing report portal. I designed, built and operate every layer of it, from the React glass interface down to the Celery beat schedule.
Production AI systems, research implementations and open-source builds across LLMs, agents and ML.
Multi-agent pharmaceutical research platform
A multi-agent platform on Google Gemini that finds new therapeutic uses for already-approved medications. Specialist agents mine clinical trial registries, patent filings and market signals, then a synthesis pass writes an evidence-based drug repurposing report with every claim traced back to its source.
Context-aware code completion for developers. Parses the surrounding file, builds a retrieval prompt from the local symbol graph and returns completions with an explanation of the change.
An agent for financial analysis and market research using LLM tool calling to fetch, reconcile and reason over live market data before writing a position summary.
Turns raw keyword lists into structured ad groups with compliant copy. Gemini 2.5 Pro primary with a GPT-4o-mini fallback through OpenRouter, exporting import-ready spreadsheets.
Neuro-AI deep learning that classifies age and gender from raw EEG brainwave signals, pairing signal preprocessing and band filtering with convolutional architectures.
Product intelligence for Amazon listings. NLP and sentiment pipelines turn thousands of reviews and market data points into ranked, actionable seller insights.
A hands-on LangChain reference covering chains, agents, RAG pipelines, memory modules and end-to-end application builds, written as runnable Python rather than snippets.
Field notes from production work: what the systems actually do, the numbers behind them, and the decisions I would make again.
How I pulled SDXL, ControlNet and IP-Adapter out of the AUTOMATIC1111 WebUI into a reproducible headless CLI pipeline with LoRA, DreamBooth and FID scoring.
ReadHow I route every AI call in Opaeron through one gateway: task-based model selection, cross-provider fallback chains, JSON repair and per-run cost tracking.
ReadHow Opaeron turns a pasted client meeting transcript into action items with owners and due dates: chunking, extraction, dedup, and a text guard.
ReadPeer-reviewed research and teaching, because shipping fast works better on strong foundations.
Hierarchical spatio-temporal video analysis. Convolutional networks handle multi-scale feature extraction, transformer language models carry semantic understanding, and a multimodal encoder-decoder with attention produces the summary.
A full session on generative AI fundamentals covering GANs and VAEs, transformer architectures, transfer learning and fine-tuning strategy, and the ethics of putting generative systems in front of real users.
Flip a card. Every award has a system behind it. (hover or tap!)
For engineering a platform-agnostic Stable Diffusion pipeline reaching 80 percent image quality accuracy while removing vendor lock-in from every internal generative workflow.
Recognized for running GPU-accelerated generative workflows across mixed hardware, driving SDXL, ControlNet and IP-Adapter from a single command line interface.
Peer-reviewed research on generative video summarization combining CNN feature extraction with transformer language models, published in the SmartCom 2025 proceedings.
Video-LLaVA-7B with Whisper large-v2, a multimedia AI benchmark at SRI-B that led directly to a Team Lead promotion.
CrewAI and LangChain agents across SEO and paid ads workflows produced 60 percent less manual work and 3x faster content delivery.
Taught GANs, VAEs, transformer architectures and AI ethics to a global webinar audience as a student researcher.
Open to forward deployed engagements, generative AI builds and interesting AI conversations. I typically reply within 24 hours.