Independent AI engineer · Available for consulting
I'm Suyash Dubey — an applied, forward-deployed AI engineer. Three years building agentic systems and retrieval pipelines in healthcare and public digital infrastructure, where a wrong output is a clinical risk rather than a bad demo.
less clinician documentation time after the scribe agent rolled out
accuracy lift on real-time EEG signal classification
manual steps removed from a content operations pipeline
shipping LLM systems in regulated, data-sensitive domains
Figures are approximate and reflect internal measurement at the time, in previous roles. Happy to walk through how each was measured on a call.
What I do
Four ways to work together. Every one of them ends in something running, instrumented and handed over — not a document about what could be built.
Zero to one. We agree the single outcome the product has to prove, then I build and deploy the thinnest system that proves it — API, model orchestration, retrieval, front end if you need one, and the evaluation that tells you whether it actually works.
Stateful LangGraph workflows with conditional branching, tool use and human-in-the-loop review, so the agent handles the messy middle instead of escalating to a person on every turn. Retrieval that is grounded in your own corpus, with the citation trail to prove it.
I sit with the people who will actually use the system, in their language, and come back with what is feasible, what it costs, what it risks and the order to build it in. Includes translating AI-governance and security requirements into concrete technical controls rather than policy prose.
The part most teams bolt on too late. Evaluation harnesses that gate releases, Langfuse tracing on every model call, versioned prompts, PII and PHI redaction before inference, and dashboards for token spend, latency and quality so regressions surface before your users find them.
How it runs
The same four steps whether it is a two-week discovery or a six-month build. The point is to get real usage in front of the system as early as it can survive it.
I talk to the people who will use it, not only the people who commissioned it, and write the behaviour spec in their words.
One end-to-end path through the system, deployed. Enough to be wrong in public and cheap to change.
Traces, evals, redaction and cost telemetry go in on the first commit — never as a pre-launch scramble.
Weekly cycles against what people actually did with it, with the evaluation suite deciding what ships.
Selected work
A LangGraph agent that listens to a physician–patient conversation, extracts clinical entities, and drafts a note grounded in patient history and clinical guidelines on AWS Bedrock — with prompt governance and full tracing, because the output touches protected health data.
Impact ≈40% less documentation timeStack LangGraph · Bedrock · Langfuse 02 EMB Global · Rayo AI platform · 2025Multi-step agents for research, drafting and optimisation behind an AI SEO platform, with brand-voice and compliance checks enforced at generation time rather than in review — orchestrated through n8n and served on FastAPI under high throughput.
Impact ≈60% fewer manual stepsStack FastAPI · LangChain · n8n · Azure 03 National Health Authority, India · ABHA · 2023The subscription management microservice for ABHA, India's national health identity platform, built from scratch — automating consent-driven subscription flows across health services, as part of Code for GovTech.
Scope National digital public goodStack Spring Boot · Java · MongoDB · DockerAgentic API that ingests a URL, scrapes it autonomously, judges it against a written policy with an LLM, and returns structured non-compliant findings with justification.
LangChain · Flask · PydanticEnd-to-end retrieval-augmented QA over unstructured documents — embedding retrieval, grounded generation and an evaluation harness, exposed as REST.
LangChain · FAISS · FastAPIZero-human-in-the-loop automation: inbound WhatsApp message via Twilio, diffusion try-on model, result delivered straight back to the sender.
Flask · Twilio · GradioEvaluated LeNet, AlexNet, VGG and GoogLeNet for in-memory computation substitution, cutting average execution time by around three minutes.
TensorFlow · Keras · SimulinkTrack record
Healthcare, public digital infrastructure, research and content platforms — consistently the environments where a wrong model output has a real cost.
LangGraph · LangChain · FastAPI · EKS · Kafka · Redis · Langfuse
PyTorch · TensorFlow · Keras · Python
Python · LangChain · LangGraph · AWS Bedrock/Lambda/S3/DynamoDB · Langfuse · TypeScript
Python · FastAPI · LangChain · n8n · GPT-4 / Claude · Azure · PostgreSQL
Spring Boot · Java · MongoDB · Docker
About
I build LLM products end to end and sit with the customer while doing it. That combination is the whole job description: most AI projects do not fail on model quality, they fail because nobody translated what the business actually needed into something a system could be held to.
My first three years were spent in places where that gap has teeth. A medical scribe running on protected health information, where an ungrounded sentence is a clinical risk. India's national health identity platform, where a consent flow has to be right for a population, not a cohort. EEG classification, where a 20% accuracy claim has to survive a benchmark against published research rather than a cherry-picked test set.
What came out of that is a fairly opinionated way of working. Ground every generation in something retrievable. Version your prompts like code. Trace every model call from day one. Put an evaluation suite between the model and your users, and let it decide what ships. Redact before you infer. None of it is exotic — it is just the difference between a demo and a system.
Today I run an independent practice, working with founders and product teams in the US, UK, Europe and Australia who need someone who can scope the problem in the room and then go and build it.
Questions
If yours is not here, ask it in the form below — I answer briefs personally.
A working system, not a slide deck. That usually means an agent or retrieval pipeline running behind an API, wired into your data, with evaluation and tracing in place so you can see what the model did and why. I scope it with your stakeholders in plain language, build the thin slice first, instrument it, then iterate against real usage.
Yes. I spent two years on a production medical scribe operating on protected health information, and contributed to India's national digital health infrastructure through Code for GovTech.
In practice that means grounding every generated output in a retrievable source, redacting PII and PHI before it reaches a model, versioning prompts, tracing every call, and agreeing the regulatory and security constraints in writing at the start of the engagement.
Three shapes. A one to two week discovery sprint that ends in an architecture and a costed roadmap. A four to six week MVP build that ends in a deployed system your users can touch. Or an ongoing fractional engagement, typically two to three days a week, embedded with your team.
LangGraph and LangChain for orchestration, FastAPI for serving, AWS Bedrock or a direct provider API for inference, FAISS or Chroma for retrieval, and Langfuse for tracing and evaluation.
I stay provider-neutral — Anthropic, OpenAI and open-weight models all have a place, and the choice should follow the evaluation results rather than the other way round.
That is most of the work I get asked for. The gaps are usually the same: no evaluation harness, no tracing, prompts living in application code, retrieval that is not really grounded, and no plan for cost or latency under load. I close those, then put the serving path on infrastructure that can carry it — Kubernetes, Kafka and Redis when the load genuinely demands them.
I work remotely from India on IST. That gives a full overlap with the APAC and Australian working day, a comfortable overlap with Europe, and US Eastern mornings. Clients in the US, UK, Europe and Australia have all worked in that rhythm.
Send a short brief through the form below, or email dsuyash57@gmail.com. If it looks like a fit, we take a 30 minute call to pressure-test the problem, and I come back with a proposed scope, a timeline and a fixed price for the first phase.
Contact
Tell me the outcome you need and the constraint you're up against. If I'm not the right person for it, I'll say so and point you at someone who is.