Building Retrieval-Augmented Systems at the Edge of Healthcare and Embedded Intelligence

1Introduction

The last three years of my undergraduate work have circled around a single observation: large language models are most useful when they are wrong least often, and they are wrong least often when they are made to retrieve rather than recall. This thesis — informal but persistent — threads through every system in this portfolio, from a kidney-pathology classifier that defers to a clinician1 to a document-chat tool that refuses to answer without a citation.

I am presently a member of the AI-Studio team at eInfochips2, where I work on retrieval-augmented pipelines for test-case synthesis — first for web applications, now for embedded device firmware. Prior to that I spent a research stint at the CSIR–Central Road Research Institute, where I built a Flask & OpenCV pipeline to measure potholes from monocular imagery [1].

This portfolio is written as a paper because that is how I think. Each section below is a real piece of work; each figure is a real system; each equation is a method I have implemented and re-implemented enough times to commit to memory.

2Current Research Programme

2.1  Retrieval-augmented test synthesis for embedded devices

Embedded test suites are notoriously brittle: they encode tacit knowledge about hardware quirks, timing tolerances, and revision-specific errata that rarely make it into formal requirements. We are building a RAG system that ingests datasheets, errata, and historical bug reports, and emits candidate test cases in the project's existing fixture vocabulary.

2.2  Earlier work: a website test-case RAG

The embedded-device system is preceded by, and benefits from, an earlier RAG pipeline I improved for web-application test generation. Lessons there — chiefly that chunking is a UX decision, not a technical one — carried over directly. Test engineers do not want a function-shaped chunk; they want a scenario-shaped one.

3Methods & Recurring Patterns

Three primitives reappear so often that I now treat them as constants of the work. They are stated below in the most compact form I can defend.

3.1  Reciprocal-Rank Fusion

Given m ranked lists from heterogeneous retrievers (dense, lexical, HyDE, …), each document d receives a fused score:

I default to k = 60, which is the value from the original TREC paper [2]. The constant matters less than the property: poorly-ranked-but-frequent documents survive, while well-ranked-but-orphaned ones do not get blindly promoted.

3.2  Confidence-gated classification

The kidney-pathology classifier in Kidnex never returns a label without an accompanying calibrated probability:

Where 𝒞 = {Normal, Cyst, Tumor, Stone}. The downstream LLM prompt is templated on p_top: below a threshold, the patient-facing copy explicitly says "this scan should be reviewed by a clinician before any action" — not as boilerplate, but as a hard control-flow branch.

def stream_answer(query, store):
    # §3.1 — fuse retrievers
    chunks = rrf([
        store.dense_search(query, k=20),
        store.bm25_search(query, k=20),
        store.dense_search(hyde(query), k=10),
    ])[:8]

    # §3.3 — every token must carry its evidence
    for tok, src in llm.stream(prompt(query, chunks)):
        yield SSE(tok, citation=src.page_anchor)

3.3  Citations as a first-class output

In Kneen, the streaming API does not emit tokens — it emits (token, source-anchor) tuples. The frontend renders each token as a clickable footnote that scrolls the original PDF to the highlighted passage. This is the single design decision I am most proud of, because it makes hallucinations visible: a missing anchor is a tell.

4Results, At a Glance

5Discussion & Where I'm Headed

The pattern is becoming hard to ignore: every time I have tried to make a model cleverer, the gains were marginal; every time I changed what it looked at, the gains were large. I would like to keep working on systems where the interesting research question is the substrate — what gets indexed, how it gets chunked, how it gets surfaced — rather than the parameter count.

Concretely, I am looking for roles in Applied AI / ML Engineering, with a preference for teams shipping production-grade RAG, retrieval, or computer-vision systems in regulated or high-stakes domains (health, automotive, embedded). My contact card is in Appendix B.