What I've Built

A.1Kneen — chat with your documents, but with receipts

What it does

Kneen is a RAG application that lets you upload documents and have context-aware, cited conversations across multiple LLM backends. Every answer comes with inline citations pointing to the exact page and text snippet in the source material — not vague references to "chapter 4", but clickable anchors that scroll the original PDF to the highlighted passage.

The thing that makes it different — Document Insights

Every time a chunk is retrieved and used to answer a question, Kneen records it. Over time this builds a per-document intelligence map: which sections have been queried, how often, and which sections have never been touched.

Users see a coverage percentage, a hotspot list of the most referenced sections, and a dead-zone count of chunks never queried. The killer feature: "Questions You Haven't Asked Yet" — Kneen analyzes the dead zones and uses the LLM to proactively suggest important questions the user never thought to ask. It turns a passive Q&A tool into an active discovery system.

How I built it

A single PostgreSQL instance with pgvectorserves as both the relational store (users, documents, sessions) and the vector store. No separate vector DB — the embedding and the chunk text live on the same primary key, so they can't drift.

Retrieval is hybrid: dense search via cosine similarity over halfvec(768), lexical search via Postgres's built-in ts_rank, both 20-deep, fused with Reciprocal-Rank Fusion. For vague queries, the system routes through HyDE first — the LLM hallucinates an answer, I embed that, and search with the hallucination's vector. The hallucination is thrown away; only its vector survives.

Streaming uses Server-Sent Events (SSE): the API emits (token, source-anchor) tuples, not bare tokens. The frontend renders each token as a clickable footnote. A missing anchor is a visible tell for hallucination — which makes the system honest by design.

What I shipped & what I learned

• 0.91 citation precision — fraction of generated claims where the cited span actually supports the claim, hand-graded on 200 query-document pairs.
• The citation is the product. Stripping out citations during a usability test cut trust scores in half — even though the answers were identical.
• "Questions you haven't asked" is a sleeper hit. Users cared more about this than any retrieval improvement that took ten times the effort.
• Single-store discipline pays off. A separate vector DB would have saved two days on the prototype and added six months of drift bugs.

A.2Kidnex — upload a CT scan, get a diagnosis and a next step

What it does

Kidnex is a kidney health platform. You upload a CT scan image and two things happen: a ResNet-50 classifier categorizes it into one of four classes — Normal, Cyst, Tumor, Stone — with a confidence score. Then that classification, along with the patient's medical history, gets passed to an LLM that generates actionable next-step suggestions. The goal: bridge the gap between a radiologist's report and a patient holding it.

Approach

A pre-trained ResNet-50 backbone is fine-tuned end-to-end on the 4-class problem using the standard cross-entropy loss with L2 regularisation:

The indicator 𝟙[y_i = c] selects the true class; the L2 term keeps the backbone from over-committing to the small training set. After training, the model outputs a top-1 label and a calibrated probability p_top. That probability controls what happens next — the LLM prompt has four conditional branches keyed on p_top's decile. 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.

The LLM never sees the scan. It sees a structured object: classification label, confidence score, and medical history. This keeps it factual — it writes the explanation, it doesn't influence the diagnosis.

What I shipped & what I learned

• 96.4% top-1 accuracy on the held-out test set. The residual 3.6% sits almost entirely on the cyst↔tumour boundary.
• The LLM's role is narrower than expected. It renders explanations; it doesn't reason about diagnoses. Treating it as a renderer, not a reasoner, was the right call.
• Calibration matters more than accuracy when the downstream consumer is a non-expert. A 96% model that lies about its confidence is worse than an 88% model that doesn't.

A.3Pothole measurement — CSIR · CRRI internship

What it does

Road authorities triage repairs by pothole area, but the current method is a person with a measuring tape. The brief from CRRI: given a single phone photo of a pothole with a reference object, return its area in cm². Target accuracy: under 15% MAPE.

How I built it

Classical CV pipeline: CLAHE→ bilateral filter → adaptive threshold → contour extraction. The reference object's known size sets the pixel-to-mm scale; pothole contour area converts from there. No GPU, no training set, no model drift — a deterministic pipeline that ran on a five-year-old laptop in the field.

def measure(img, ref_obj_mm):
    g  = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    g  = clahe.apply(g)
    g  = cv2.bilateralFilter(g, 9, 75, 75)
    th = cv2.adaptiveThreshold(g, 255, ADAPTIVE_GAUSSIAN, ...)

    cnts, _ = cv2.findContours(th, RETR_EXTERNAL, CHAIN_APPROX_SIMPLE)
    ref     = largest_rectangular(cnts)         # the marker
    pothole = largest_irregular(cnts, exclude=ref)

    px_per_mm = perimeter(ref) / (4 * ref_obj_mm)
    return cv2.contourArea(pothole) / (px_per_mm ** 2)

What I shipped & what I learned

• 8.7% MAPEon a hand-measured set of 120 potholes — the brief required <15%.
• Classical CV is underrated for narrow problems. No GPU, no training data, no model versioning — just a deterministic pipeline that actually ships.
• Lighting is the entire problem. Half the engineering time went into the CLAHE + bilateral combination; the contour step was an afternoon.

A.4What I'm building now at AI-Studio

Two systems under active development at eInfochips · AI-Studio. NDAs limit detail — high-level descriptions only.