Portfolio

Case studies packaged around real business proof

We are not showing feature lists for their own sake. Each project is framed the way a serious buyer evaluates delivery: the business problem, the constraints, the intervention, the outcome, and the operational artifacts that made it usable in the real workflow.

Request estimate View work

Problem first

We start with the blocked workflow, not a generic AI capability pitch.

Constraints included

Privacy, latency, noisy inputs, messy data, and deployment realities are part of the case study.

Artifacts shipped

Every engagement leaves behind usable systems, handoff material, and production-ready assets.

RAG systems

RAG-Powered Chatbot

A document assistant built for teams that needed reliable answers from internal PDFs without forcing people to dig through folders, repeat support questions, or lose conversation context.

Problem

Internal knowledge was trapped in PDFs, making answers slow, inconsistent, and hard to reuse.

Constraints

The system needed fast retrieval, secure document handling, and persistent chat context for auditability.

Intervention

We built PDF ingestion, Gemini embeddings, FAISS retrieval, and a Django application packaged for Cloud Run deployment.

Outcome

Teams got faster self-serve answers, more consistent responses, and a support workflow that can be improved over time.

Operational artifacts

PDF ingestion and storage workflow
Vector retrieval pipeline
Chat history for continuity and auditability
Containerized deployment for Cloud Run

Stack

Django
Gemini
FAISS
PostgreSQL
Docker
Cloud Run

Speech intelligence

Voice Intel

A conversational intelligence workflow designed to reduce documentation load, capture the important parts of real conversations, and turn audio into usable follow-up actions.

Problem

Teams were spending too much time turning calls and visits into notes, summaries, and next steps.

Constraints

The workflow had to handle real speech, multiple speakers, and operational follow-up items without adding more admin work.

Intervention

We built audio capture, transcription, speaker differentiation, and structured note generation with follow-up detection.

Outcome

Documentation became faster and more consistent, with clearer handoffs on labs, referrals, and next actions.

Operational artifacts

Audio capture workflow for live calls
Transcript and speaker separation pipeline
Structured summary and note templates
Follow-up detection for action tracking

Stack

Django
PostgreSQL
Speech-to-Text
NLP

Computer vision

Image Semantic Search

A visual discovery workflow for situations where users know what they want to find, but cannot describe it well enough for keyword search to work.

Problem

Product discovery broke down when users could only show an image, not type the right words.

Constraints

Search needed to feel immediate, work on mobile and desktop, and scale beyond a single demo dataset.

Intervention

We used image embeddings and similarity ranking to turn uploaded visuals into retrieval-ready queries.

Outcome

Users could move from “I have a picture” to relevant product matches without relying on exact text descriptions.

Operational artifacts

Image embedding pipeline
Similarity ranking service
Upload and search interface
Scalable backend patterns for retrieval

Stack

Python
Django
VGG19
Embeddings

Infrastructure AI

Road Construction Site Analyzer

A satellite analysis workflow for infrastructure teams that needed better road-surface estimates from imagery that was messy, partially hidden, and hard to segment cleanly.

Problem

Manual and naive image analysis struggled to estimate road area accurately from satellite imagery.

Constraints

Shadows, foliage, and partial occlusions made standard segmentation unreliable in the hardest scenes.

Intervention

We combined SAM2 segmentation with multimodal reasoning to refine visible masks and recover hidden road coverage.

Outcome

Area detection error dropped by 33.7%, giving the client a much more reliable measurement workflow.

Operational artifacts

High-resolution segmentation workflow
Mask refinement for shaded and covered areas
Visual QA outputs for review
Repeatable measurement pipeline

Stack

SAM2
Multimodal AI
Satellite imagery

Real estate AI

SceneStatte.ai

A media-generation workflow for property teams that needed to turn raw listing assets into polished, repeatable virtual tour experiences without building each story by hand.

Problem

Real estate teams needed listing content that felt premium without slowing down every property launch.

Constraints

The workflow had to turn raw media and property data into something repeatable enough for real teams to use at scale.

Intervention

We built a GenAI media pipeline for automated tour generation, highlight selection, and spatial storytelling.

Outcome

Property content production became faster and more consistent, with a workflow better suited to repeated listing operations.

Operational artifacts

Virtual tour generation workflow
Highlight selection and sequencing
Export-ready property showcase assets
Team-friendly workflow integration

Stack

GenAI
Media pipelines
GCP
Gemini API
Docker

Education AI

ChildBook AI

A personalized storytelling workflow that turns a child’s interests into a finished narrative and matching illustrations without requiring manual creative production for every book.

Problem

Creating custom children’s books one at a time is slow, expensive, and difficult to scale.

Constraints

The experience had to keep story and imagery aligned while staying simple enough for end users to guide.

Intervention

We built a workflow for personalized narrative generation, illustration generation, and export-ready story packaging.

Outcome

The product turned a manual creative process into a repeatable personalized experience for families.

Operational artifacts

Prompt orchestration for custom stories
Illustration generation workflow
Personalized inputs and story controls
Digital and print-ready export flow

Stack

Gemini API
Story generation
Image generation

Healthcare AI

An intelligent documentation and support workflow for clinical environments where time is limited, patient context is fragmented, and information needs to be structured enough to use safely.

Problem

Providers were losing time to note-taking and fragmented patient data instead of focusing on care.

Constraints

The workflow required secure handling patterns, structured outputs, and usable support inside sensitive clinical contexts.

Intervention

We combined transcription, multimodal extraction, documentation support, and trend analysis in one operational workflow.

Outcome

Clinical teams got less administrative overhead and more structured patient information at the point of use.

Operational artifacts

Automated note-generation workflow
Secure processing and storage patterns
Trend and insight extraction
Structured outputs for clinical review

Stack

Multimodal AI
FastAPI
PyTorch
GCP