From pixels and language to systems people can use.
I’m an AI/ML engineer working across computer vision, intelligent document workflows and applied AI. My work connects medical-device label inspection, manufacturing vision, conversational tools and research on 3D medical images.
Computer vision · Applied AI · Research
Overview
ENGINEERING CAPABILITIES
Research fluency. Production instinct.
I take AI from experiment to deployment. I build and evaluate models, connect them to the services people use, and use operator feedback and delivery traces to improve the workflow.
End-to-end ownership
From prototype to production.
Design & prototype
Train & evaluate
Deploy & operate
Trace & improve
01 / Deployed systems
Production AI & deployment
I led ASTRA’s OCR upgrade and contributed to the AWS migration. The workflow connects label capture and extraction to operator review, inspection storage and background delivery. Delivery tracking helps distinguish a saved inspection from a record successfully received downstream.
QuoteMaster’s prototype searches the approved catalog, queues room previews and generates estimates from confirmed inputs. The agent workflow under development adds persistent state, human confirmation and retry handling. Catalog IDs and pricing rules stay in application logic.
I work on retrieval where the source matters. QuoteMaster retrieves from approved catalog and FAQ text. In ASTRA, extracted fields lead back to their source image, so operators can inspect the evidence for a value before confirming it.
Catalog and FAQ retrievaldocument OCRevidence mapping
I work across industrial inspection and medical imaging. Inspect AI compares a work area against a known-clear reference and highlights changes for review. My published lung-nodule research tests multi-scale 3D context, attention and model fusion for reducing false-positive candidates.
A predicted label field is more useful when an operator can see the image region behind it. In ASTRA, extraction and review belong in the same workflow.
⌁
Measure the right boundary
Inference time, end-to-end latency and operator time answer different questions. I keep the dataset, measurement and limits beside each result.
◎
Design around the person
Line clearance remains an operator decision. Vision highlights changes, and feedback helps refine the next inspection.
✧
Connect research to implementation
My lung-nodule work uses local and surrounding 3D context to reduce false-positive candidates. It also taught me to make evaluation splits and model fusion explicit.
PROJECTSselected systems and the evidence behind them
ASTRA · Intelligent label verification
Package photos become extracted identifiers, barcode values and symbol detections that an operator can review against their visual source.
PaddleOCRPythoncomputer visionFastAPIAWS
ASTRA · Intelligent label verification
PaddleOCRPythoncomputer visionFastAPIAWS
Computer vision · OCR · Medical devices
Package photos become extracted identifiers, barcode values and symbol detections that an operator can review against their visual source.
My contribution
On a Johnson & Johnson client engagement, I led the PaddleOCR v5 upgrade and worked on the AWS migration associated with DePuy Synthes. The system combines label capture, extraction, review, local inspection storage and downstream delivery.
How it works
Capture outer and sterile label images → run OCR, barcode and symbol recognition → map extracted fields to evidence → operator review → save the inspection → deliver records through background workers.
Separate delivery paths matter. A saved tablet record and a successfully received downstream record are different events, so delivery status must be traceable independently.
My current résumé reports label verification falling from 12 minutes to under one minute, with 400 hours saved per month. These are reported operational outcomes; the portfolio does not present them as an independently reproduced benchmark.
Computer vision · Manufacturing · Proof of concept
A camera compares an inspection area with a known-clear reference and highlights changes for an operator to review before the next run.
My contribution
I enhanced the hybrid vision pipeline for the Schaffhausen manufacturing proof of concept. Image differencing and structural comparison find candidate changes; ResNet feature comparisons help inspect the relevant regions. Operator feedback supports reviewed threshold updates.
Evaluation
The July 2026 test records contain 297 correct runs out of 300, where a correct run has no recorded false positive or false negative. The set combines 150 physical-test runs and 150 simulated-video runs.
Mean recorded inference time was 0.191 seconds. This measures model processing, not capture-to-screen latency. The physical subset had 147 correct runs out of 150. These prototype results do not establish performance on every plant or camera.
QuoteMaster · From product questions to a draft quote
A design and quoting assistant connects approved product information, room previews and deterministic price calculations for curtains, blinds and related interior products.
Applied AI · Retrieval · Multimodal product workflow
A design and quoting assistant connects approved product information, room previews and deterministic price calculations for curtains, blinds and related interior products.
Current implementation
The documented Django prototype searches approved catalog and FAQ text, creates jobs for illustrative room previews and renders a PDF estimate from confirmed product selections and measurements. Company-scoped records and explicit pricing rules keep commercial data separate from generated imagery.
Agent workflow under development
The engineering plan extends this into a Telegram and LangGraph workflow with persistent state, customer confirmation, tenant-scoped hybrid retrieval and retry-aware media jobs. A proposed evaluation measures time to a correct quote, evidence retrieval, citation support and room-region mask quality.
Engineering choices
The model can help interpret a request; product IDs and price rules come from the approved catalog. A generated room image is an illustration, not a measurement or guarantee of installed appearance. Unsupported custom products route to manual estimation.
Current tools: Django, Python, approved-text retrieval, OpenCV GrabCut, image-editing integration, PDF generation.
Small nodules and surrounding anatomy can look similar in CT. Our research scores candidates from an upstream detector using both local detail and wider 3D context.
MP-AFNet · Lung-nodule false-positive reduction
3D computer vision · Medical imaging · Published research
Small nodules and surrounding anatomy can look similar in CT. Our research scores candidates from an upstream detector using both local detail and wider 3D context.
My contribution
I am the first author of a study combining multi-patch 3D convolutional networks, attention and hybrid model fusion. My research assistant work at BML Munjal University focused on this candidate-classification stage.
Method
Extract 16³, 32³ and 48³ voxel patches around each candidate. Custom, residual and dense attention models use these scales to score the candidate. Hybrid fusion combines the models that agree with the majority decision.
Result and scope
The paper reports 0.931 CPM on LUNA16 with ten-fold cross-validation. CPM is mean sensitivity at seven false-positive rates per scan. This is a paper-reported research result, not a clinical validation claim or a newly reproduced score from the later PyTorch implementation.
A Jetson Nano camera pipeline supplies an inbound vehicle count and supporting snapshot for comparison with RFID observations.
My contribution
I owned the computer vision pipeline, including vehicle detection, multi-object tracking and a calibrated passage-counting rule. The archived MRV implementation uses YOLOv4-tiny and FastMOT. The wider résumé also describes YOLOv5 and Deep SORT work; those should not be conflated with this specific snapshot.
How it works
Camera frames → vehicle detections → stable tracks → directional crossing rule → count and supporting image → downstream reconciliation.
A count discrepancy is a reason to review the evidence. Occlusion, missed RFID reads and timing differences can all produce one.
Dealer self-service · Finance information in WhatsApp
Mahindra dealers can request finance information, including outstanding balances, receipts and ageing reports, through a familiar messaging interface.
RasaInfobipFlaskRedisSQL Server
Dealer self-service · Finance information in WhatsApp
RasaInfobipFlaskRedisSQL Server
Conversational AI · NLP · Enterprise integration
Mahindra dealers can request finance information, including outstanding balances, receipts and ageing reports, through a familiar messaging interface.
My contribution
I trained the Rasa language-understanding model and integrated the Infobip connector. Conversation state and detected intent route requests to menus, data lookups or document responses. The SAP team owned the upstream finance-data integration.
Impact
Infobip’s customer story reports a 30% increase in operational efficiency for the overall WhatsApp self-service rollout. This is a rollout-level result, not a separately measured uplift attributable to one component.
EXPERIENCEwhere I have worked, and what I built there
May 2025–Present
Johnson & Johnson · AI Engineer
May 2025–Present · United States · Client engagement
Computer vision and intelligent inspection work across medical-device label verification, line clearance, annotation and PPE prototypes. Led the ASTRA OCR upgrade and worked on migration and downstream record reconciliation.
Worked on AI/ML software and product workflows. The QuoteMaster case study presents the currently documented prototype and its agentic development roadmap separately.
Developed computer vision and language-understanding components for vehicle tracking, plate recognition, industrial inspection, camera-based monitoring and dealer self-service.
Attention Based Multi-Patched 3D-CNNs with Hybrid Fusion Architecture for Reducing False Positives during Lung Nodule Detection
Vamsi Krishna Vipparla, Premith Kumar Chilukuri and Giri Babu Kande. Journal of Computer and Communications, 9(4), 1–26, 2021. DOI: 10.4236/jcc.2021.94001.
The work investigates how multi-scale 3D context, attention and model agreement can reduce false positives among lung-nodule candidates. The paper reports 0.931 CPM on LUNA16.
5 indexed citations, checked November 15, 2026 in OpenAlex, Semantic Scholar and Crossref. Citation counts depend on the index and can change. Citations indicate scholarly use; they do not establish clinical adoption.
Business questions and actions become cards and task modules inside Microsoft Teams.
Microsoft TeamsAzure BotAdaptive CardsNode.jsReact
Teams · Business workflows
Mahindra · project archive
Microsoft TeamsAzure BotAdaptive CardsNode.jsReact
The archived team project connects Teams activities to Dexter and HR analytics backends. Separate message and submit handlers preserve conversation context; Adaptive Cards and task modules return the result. The source establishes the integration design, without a measured adoption claim.
Mahindra · project archive
Phone sensor acquisition
A phone becomes a source of three-axis accelerometer readings for a local visualization.
PythonSensorDroidSensor callbacks
Phone sensor acquisition
Mahindra · project archive
PythonSensorDroidSensor callbacks
The prototype discovers a SensorDroid device, subscribes to acceleration callbacks and exposes the latest sample to plotting scripts. It demonstrates acquisition and visualization, rather than a calibrated motion classifier. This is a project-archive example; personal implementation ownership is not established.
Mahindra
ANPR · Gate-camera plate recognition
Motion-triggered vehicle detection isolates high-resolution plate images for OCR and gate records.
YOLOOpenCVGoogle Cloud VisionFlaskMySQL
ANPR · Gate-camera plate recognition
Mahindra
YOLOOpenCVGoogle Cloud VisionFlaskMySQL
My contribution covered camera detection regions, motion triggers, YOLO vehicle and plate detection, and mapping plate crops back to the high-resolution frame. The wider system provides OCR, vehicle logs and operator review. A later archive adds Deep SORT; these are distinct implementation snapshots. The résumé also describes boom-barrier integration, while archive evidence does not establish a current rollout status.
Mahindra · project archive
Antyodaya · PDF table ingestion
State documents move through discovery, download, table extraction and database preparation.
PythonTabulaDockerPDF extraction
Antyodaya · PDF table ingestion
Mahindra · project archive
PythonTabulaDockerPDF extraction
The archived team workflow separates a manager scraper from containerized parsing workers. State files track work and retries; Tabula extracts table CSVs before normalization. Document completion and field correctness require separate evaluation, so no unmeasured processing savings are claimed.
Mahindra · project archive
Basebot · HR conversational services
Employee questions route to policies, FAQ retrieval or HR actions within a modular chatbot.
RasaDjangoFastAPIRedisWhooshGensim
Basebot · HR conversational services
Mahindra · project archive
RasaDjangoFastAPIRedisWhooshGensim
The archived Mahindra HR proof of concept combines a Django gateway, Rasa workers, response generation, policy search and custom actions. A Teams adapter handles ticket and time-off workflows separately. This describes the team archive rather than attributing the full platform to one engineer; live integration results are not established.
Mahindra
Tractor decal verification
Locate the tractor and decal region before checking model-identifying text on both sides.
YOLOv3-tinyOpenCV DNNOCR
Tractor decal verification
Mahindra
YOLOv3-tinyOpenCV DNNOCR
I trained the YOLO detection models that locate tractors and decal regions. The surrounding workflow reads text, compares it with the expected tractor variant and presents inspected images to an operator. The archive supports the detector and workflow; the résumé’s manual-effort reduction has no retained matched-baseline evaluation.
Mahindra · project archive
Engine test analytics
A dashboard connects test measurements with parameter definitions and tolerance limits.
Data extractionAnalytics APIsDashboards
Engine test analytics
Mahindra · project archive
Data extractionAnalytics APIsDashboards
The archived team project separates test-data extraction from parameter and analytics services. Engineers can inspect filtered readings and distributions against specification boundaries. Units, program definitions and selected populations must agree before interpreting differences. No measured defect reduction is established.
Mahindra
Face localization API
A still-image service returns face boxes and confidence scores for downstream processing.
OpenCV DNNCaffe SSDFlask
Face localization API
Mahindra
OpenCV DNNCaffe SSDFlask
My responsibilities included computer vision methods. This specific prototype uses a bundled Res10 SSD checkpoint through OpenCV DNN and returns detections above a supplied threshold. Despite an endpoint named face_recognition, this service does not return identities or embeddings; checkpoint training provenance is not established.
Mahindra
Face-mask detection
A two-class detector identifies faces and mask status from an uploaded image.
YOLO tinyOpenCV DNNFlask
Face-mask detection
Mahindra
YOLO tinyOpenCV DNNFlask
I worked on the computer vision method and model training. The archived implementation performs a single detector pass over the image and aggregates face detections into an API response. Training records and held-out evaluation would be needed to associate a particular checkpoint with a measured accuracy claim.
Mahindra
ArcFace · Face recognition
Enrollment and recognition use face embeddings within a subscription-scoped gallery.
ArcFaceRetinaFaceFlaskMongoDBDocker
ArcFace · Face recognition
Mahindra
ArcFaceRetinaFaceFlaskMongoDBDocker
My contribution covered the ArcFace computer vision method and model training. RetinaFace supplies faces and landmarks, and ArcFace embeddings support nearest-gallery matching. The archived services separate registration and recognition. Personal training-run lineage and measured recognition performance are not established by the snapshot alone.
Mahindra · project archive
Interactive camera demo
A desktop event prototype renders face detections and classifier outputs from a camera feed.
OpenCVMTCNNTkinterPython
Interactive camera demo
Mahindra · project archive
OpenCVMTCNNTkinterPython
The team archive connects camera capture, MTCNN face detection, an emotion-classification path and a Tkinter display with logging. Emotion labels are model outputs, not reliable measurements of a person’s internal feelings. The archive establishes a demonstration, without measured recognition performance.
Mahindra · project archive
Supplier allocation planning
Demand, supplier mappings and logistics costs feed a transportation allocation workflow.
PythonFlaskOperations researchSQL
Supplier allocation planning
Mahindra · project archive
PythonFlaskOperations researchSQL
The archived team application prepares demand and invokes a transportation allocation routine, including the North West Corner method. That method provides an initial feasible solution under its assumptions; it does not by itself prove optimality. The case is about constraints and traceable allocations, not an unsupported cost-saving claim.
Mahindra · project archive
PAN card extraction
Card localization reduces unrelated page text before OCR and structured-field parsing.
YOLOFastAPIGoogle Cloud VisionPDF processing
PAN card extraction
Mahindra · project archive
YOLOFastAPIGoogle Cloud VisionPDF processing
The archived document workflow validates uploaded images or PDFs, localizes and crops the PAN card, applies OCR and maps recognized text into fields. Localization and exact field correctness are separate concerns. This is team project context; no held-out accuracy or individual end-to-end ownership is claimed.
Mahindra · project archive
Partner API request signing
A compact integration turns a serialized request into an HMAC authorization header.
PythonHMAC SHA-256REST
Partner API request signing
Mahindra · project archive
PythonHMAC SHA-256REST
The archived integration signs the exact request method, URI, serialized body and nonce. It illustrates why byte-level agreement matters even when two JSON objects contain equivalent data. The source establishes the signing algorithm; partner acceptance and production deployment are not established. No operational credentials are published.
Mahindra · project archive
Sajhedaari · Multilingual voice workflow
Speech and text questions route through language understanding, FAQ retrieval and custom actions.
RasaDjangoSpeech interfacesDocker
Sajhedaari · Multilingual voice workflow
Mahindra · project archive
RasaDjangoSpeech interfacesDocker
The archived team project contains English, Hindi, Marathi and Telugu text and speech assets. Transcription, intent selection, retrieval and response delivery are separate stages, with several search variants represented in the archive. No unsupported end-to-end accuracy figure or sole-ownership claim is attached.
Mahindra
Sealant bead inspection
Inspect visible sealant continuity and highlight suspect regions before assembly.
OpenCVU-NetSSIMImage registration
Sealant bead inspection
Mahindra
OpenCVU-NetSSIMImage registration
My CV contribution focused on integrating U-Net pixel-level segmentation into the vision architecture. The factory baseline used classical OpenCV on Raspberry Pi hardware managed by the hardware team. The final U-Net deployment and measured uplift are not established. Visual bead inspection does not measure seal strength or leak-tightness.
Mahindra
Camera geometry · Proximity monitoring
Person detection and calibrated ground-plane geometry turn fixed-camera views into proximity indicators.
YOLOv5mHomographyOpenCVFlaskDBSCAN
Camera geometry · Proximity monitoring
Mahindra
YOLOv5mHomographyOpenCVFlaskDBSCAN
I owned the vision portion: person detection, four-point homography calibration and per-view proximity thresholds. The wider system samples camera frames and can produce overlays, crowd snapshots and periodic reports. Calibration and perspective assumptions matter; there is no unsupported universal distance-accuracy claim.
Mahindra · project archive
SOP viewer · Scan to procedure
A QR code connects a work location with its procedure PDF in an Android viewer.
AndroidJavaPythonAmazon S3QR codes
SOP viewer · Scan to procedure
Mahindra · project archive
AndroidJavaPythonAmazon S3QR codes
The archived team workflow uploads SOP PDFs, generates QR codes from their URLs and displays the retrieved PDF after a mobile scan. The important boundary is document identity and revision management: successful scanning alone does not establish that the intended procedure was retrieved.
Mahindra · project archive
Genie · Employee-service API adapter
A Django adapter reshapes upstream employee-service data for a Teams application.
DjangoRESTGoogle App Engine
Genie · Employee-service API adapter
Mahindra · project archive
DjangoRESTGoogle App Engine
The archived team service provides routes for employee details, calendars, paystubs and related records. Its role is contract adaptation between the Teams client and upstream APIs. Source and handoff establish the interface, without claiming measured usage or reliability.
Mahindra · project archive
International Knowledge Portal
A reader portal and separate CMS make company and country information accessible and maintainable.
ReactDjango RESTMySQLCeleryRedis
International Knowledge Portal
Mahindra · project archive
ReactDjango RESTMySQLCeleryRedis
The Mahindra team archive separates the strategy portal, editor CMS and Django API. Background jobs validate spreadsheet imports and update the data used by portal views. This case illustrates controlled publishing and ingestion; individual ownership and measured adoption are not established.
Mahindra · project archive
Asynchronous invoice extraction
Uploaded invoices pass through template or QR processing before structured results are returned.
The archived team pipeline decouples upload from scheduled extraction. Text-based templates and QR decoding provide separate routes to invoice fields. Stored inputs, notifications and results make the handoffs traceable. Historical worker scheduling is configuration rather than an observed latency result.
BML Munjal University · Research
Lung nodule research · PyTorch implementation
An implementation of multi-scale attention CNNs makes candidate scoring and FROC evaluation explicit.
PyTorch3D CNNsAttentionLUNA16FROC
Lung nodule research · PyTorch implementation
BML Munjal University · Research
PyTorch3D CNNsAttentionLUNA16FROC
This implementation relates to my coauthored lung-nodule study. It resamples CT, extracts centered 16³, 32³ and 48³ patches, scores attention models and applies hybrid fusion. Local pipeline checks and the paper’s 0.931 CPM result are different evidence: a full LUNA16 reproduction was pending in the archived notes.
A desktop workflow flattens interactive PDFs for downstream processing while retaining originals.
Robin RPAAdobe PDFWindows Task Scheduler
Robin · Interactive PDF conversion
Mahindra · project archive
Robin RPAAdobe PDFWindows Task Scheduler
The archived team automation prints XFA documents to PDF through Adobe, writes conversions to a pickup directory and archives source files. It demonstrates a desktop integration boundary; a scheduled invocation is not proof of successful conversion. No historical completion rate is claimed.
Mahindra · project archive
Mahindra Rise · Guided site navigation
A website assistant turns menu selections into the next action or relevant destination.
AWS LambdaAPI GatewayDynamoDBFlask
Mahindra Rise · Guided site navigation
Mahindra · project archive
AWS LambdaAPI GatewayDynamoDBFlask
The handoff describes an AWS deployment with request handling and transaction logging; an earlier Flask implementation is also retained. The archived team project is a guided navigation workflow, not a generative AI assistant. Response correctness and analytics delivery are distinct checks.
Mahindra · project archive
MSolve · Guided desktop remedies
Verified support procedures become signed desktop tools offered through an employee-support workflow.
PowerShellPS2EXECloud FunctionsDatastore
MSolve · Guided desktop remedies
Mahindra · project archive
PowerShellPS2EXECloud FunctionsDatastore
The archived team project packages PowerShell remedies as signed executables and records execution through a logger. It connects a documented SOP to a downloadable support action. A successful script launch is not automatically a successful resolution; no support-volume reduction is claimed.
Mahindra · project archive
Vehicle telemetry delivery
A portal turns a vehicle, date range and parameter selection into a downloadable CSV.
ParquetBigQueryCloud StorageDatastoreApp Engine
Vehicle telemetry delivery
Mahindra · project archive
ParquetBigQueryCloud StorageDatastoreApp Engine
The archived team workflow finds relevant Parquet files, projects selected columns, merges records and sends an export link by email. Keeping preparation separate from delivery supports long-running requests. The case does not assert an unmeasured export speed or deployment scale.
Mahindra · project archive
MWatcher · Website health monitoring
Regional probes monitor website health alongside certificate and domain expiry.
AWS LambdaAmazon SQSMonitoringAlerting
MWatcher · Website health monitoring
Mahindra · project archive
AWS LambdaAmazon SQSMonitoringAlerting
The archived team design schedules URL checks through queues, invokes regional Lambda probes and collects results for dashboards and alerts. It distinguishes regional unavailability from missing monitoring jobs and expiry risks. No measured uptime or alerting SLA is established.
Mahindra · project archive
Supplier test-certificate extraction
Certificate images and QR regions become structured records for supplier and admin portals.
Google Cloud Visioninvoice2dataPub/SubDatastorepyzbar
Supplier test-certificate extraction
Mahindra · project archive
Google Cloud Visioninvoice2dataPub/SubDatastorepyzbar
The archived team workflow connects portal uploads, object storage, scheduled OCR and persisted results. Template extraction and QR localization/decoding serve different layouts. The source establishes the processing chain, while labeled field accuracy and personal end-to-end ownership remain unverified.
Mahindra · project archive
Visiting-card contact extraction
A card image becomes structured contact fields through provider and local OCR alternatives.
FlaskMicrosoft Form RecognizerTesseractAddress parsing
Visiting-card contact extraction
Mahindra · project archive
FlaskMicrosoft Form RecognizerTesseractAddress parsing
The archived team API supports a Microsoft recognition path and a local Tesseract path, with parsing and optional geocoding for addresses. Recognized text and enriched geographic fields are different outputs. No unsupported accuracy, latency or cost advantage is claimed.
Johnson & Johnson · client engagement
CVAT + SAM · Interactive annotation
Promptable segmentation proposes editable masks while annotators retain control of the labels.
CVATSAMNuclioDjangoPostgreSQLDocker
CVAT + SAM · Interactive annotation
Johnson & Johnson · client engagement
CVATSAMNuclioDjangoPostgreSQLDocker
I upgraded an internal CVAT platform and integrated SAM for precise crystallization segmentation. The test architecture connects the model runtime, annotation UI and organization access policies. The résumé reports a 60% acceleration; the available archive does not retain the timing protocol needed to reproduce that figure.
Johnson & Johnson · client engagement
Gowning · Visual PPE inspection
Reference matching and garment-specific checks display visible PPE evidence for review.
SAMResNet50MTCNNYOLOStreamlit
Gowning · Visual PPE inspection
Johnson & Johnson · client engagement
SAMResNet50MTCNNYOLOStreamlit
The prototype combines segmentation proposals, reference-image embeddings, face localization and specialized garment checks, then presents overlays and structured output. The current résumé places this work on a Johnson & Johnson client engagement. The archive establishes the prototype, but not a labeled evaluation validating the résumé’s accuracy figures.
Mahindra
Tractor chain and tyre alignment
Object detection supports inspection of chain and tyre alignment on industrial tractors.
UltralyticsObject detectionComputer vision
Tractor chain and tyre alignment
Mahindra
UltralyticsObject detectionComputer vision
The current résumé describes automated tractor chain and tyre alignment using Ultralytics object detection. No corresponding detailed technical reference, dataset, evaluation or public repository was identified in the supplied archive, so this remains a résumé-level project summary without additional performance claims.
TOOLKITtools I have worked with
Computer vision
PyTorchTensorFlowOpenCVYOLOSAMPaddleOCR3D CNNstracking and image segmentation
Language and AI applications
RasaOpenAI APIsLangChainretrieval workflows and agent design
My path began with medical-imaging research and continued through automotive manufacturing, AI/ML software and medical-device inspection. I enjoy problems where models have to work with physical constraints, incomplete inputs and real operational handoffs.
I bring experience in Python, model development and system integration, and I’m interested in roles that connect computer vision and agentic AI with useful, measurable products.
LET’S CONNECT
If you’re building applied AI, computer vision or ML systems, I’d like to hear about the problem your team is working on.