Deduct AI

Field officers need fast, defensible suggestions for which legal sections apply to a scenario, plus a way to surface relevant evidence, without waiting on a slow manual legal search.

A retrieval-augmented assistant that indexes legal text in a vector store and pairs it with a computer-vision pipeline for evidence detection, exposed through a lightweight field-team app.

Built the query-retrieval system over the legal-text vector store, trained the YOLO-based evidence detector on hand-labelled data, and deployed the inference endpoints.

Retrieval-augmented generation over a legal-text vector store, with an LLaMA-13B model producing section suggestions grounded in retrieved passages.

Grounding a generative model in a retrieval layer made its suggestions far more reliable than prompting alone, and cheap enough to run on a modest AWS box.

• Had to run on modest, low-cost cloud infrastructure (AWS T2-micro) rather than GPU-backed inference servers.
• Needed a field-usable client, not just a research notebook.
• Legal suggestions needed to be traceable back to source text, not free-floating generative claims.

Split the problem in two: a retrieval-augmented language pipeline for legal-section suggestions grounded in an indexed corpus, and a separate computer-vision pipeline for evidence detection, unified behind one FastAPI service and a Flutter client.

The Flutter app collects a short scenario description and optional evidence photos, then submits both to the API and renders section suggestions and detected objects in a simple field-friendly layout.

FastAPI receives the request, forwards text to the LangChain retrieval pipeline and images to the YOLO endpoint, then merges both results into a single response.

The scenario text is embedded and matched against a Chroma-indexed corpus of legal text. The top passages are passed as context to LLaMA-13B, which produces a suggestion grounded in that retrieved text rather than relying on parametric memory alone. Evidence photos are handled separately by a YOLO model trained on hand-labelled data with LabelImg.

• Chose retrieval-augmented generation over fine-tuning so the legal corpus could be updated without retraining the model.
• Kept vision and language pipelines independent so either could be swapped without touching the other.
• Used FastAPI for the gateway to get async I/O and simple request validation with minimal ceremony.

• Keeping inference latency reasonable on a T2-micro instance meant being deliberate about model size and batching.
• Hand-labelling enough evidence photos for a usable detector was the slowest part of the project.

• A 13B parameter model is smaller and cheaper to host than larger alternatives, at some cost to suggestion nuance, which is a fair trade since the retrieval grounding does most of the heavy lifting anyway.
• Optimized for a working field prototype over a fully productionized, horizontally-scaled service.

The prototype placed 3rd at the Vimarsh 5G national finals, judged by the Department of Telecommunications and IIT Madras.

• Grounding generation in retrieval is a much bigger reliability lever than prompt tweaking alone.
• Separating concerns between the vision and language pipelines made the system far easier to debug under deadline pressure.

• Add an evaluation harness to measure suggestion quality against a labelled set of scenarios.
• Explore swapping the fixed retrieval step for a lightweight agent that can decide when to search versus answer directly.