Abhay A.
I build AI systems that hold up outside the notebook — agentic RPA that replaces costly bot licences, MCP tool servers and LLM gateways that let agents act inside real enterprise software, retrieval-driven language systems, and computer vision at the edge. I start with the business case — process discovery, scope and ROI — then architect, evaluate and ship. Less demo, more uptime.
Working notes on
how I build.
01 / ABOUT
My start in AI wasn't about training models — it was frustration watching good research fail in the real world because it was too slow, too expensive, or too fragile to trust. At Safetywhat, an early-stage startup, I learned how systems actually behave under pressure: what breaks at 2am, what a client actually needs versus what a paper promises.
That taught me where real innovation happens — at the constraints. How do you run 3D posture estimation on a device with no GPU? How do you fit a large language model into a serverless deployment that costs pennies instead of dollars an hour? More recently: how do you let an AI agent safely take action inside a company's actual ERP, instead of just talking about it?
CPU-first computer vision and knowledge distillation, built for hardware that has real limits.
RAG pipelines and multi-agent systems that stay grounded, fine-tuned rather than prompted-and-hoped.
Agentic RPA on open-source computer-use tooling (Cua), MCP tool servers and LLM gateways that connect agents to ERP and business systems — without per-bot licensing.
Business-analyst groundwork first: mapping processes, finding automation gaps, and sizing scope and ROI before anything gets built.
AI evals and LangSmith tracing to measure agent accuracy and catch hallucinated tool calls before they reach production.
Client work stays private, but I build in the open too — multi-agent tooling, vision-language research and real-time systems are on my GitHub. Either way, the goal is the same: end-to-end systems that turn hard constraints into an advantage.
Work history.
02 / EXPERIENCE— PRESENT
● current
- Leading AI and automation discovery for a large enterprise as AI business analyst: assessed 600+ processes and shortlisted 100+ process cycles for AI and automation rollout, defining scope and ROI for each.
- Designing an agentic RPA platform that pairs custom LLM agents with a plugin on the open-source Cua computer-use framework — automating repetitive work across departments while cutting the per-bot licensing cost of tools like UiPath.
- Architecting MCP tool servers that expose ERP and CRM functions to agents behind an LLM gateway (model routing, fallback, cost control), with human-in-the-loop approvals for high-risk actions. Platform onboarding is underway.
- Defining the AI evals and LangSmith tracing plan to verify agent accuracy and catch hallucinated tool calls before rollout.
— APR 2026
- Re-engineered LLM deployments using 4-bit quantization and RunPod serverless infrastructure, cutting LLM inference cost by ~80%.
- Architected a 3-stage fine-tuned LLM pipeline with 4-bit quantization, achieving 98% high-quality contextual outputs.
- Built a fine-tuned AI review-management system that generates rating-aware, human-sounding responses at scale.
- Shipped multilingual chat solutions on Meta's NLLB framework paired with autonomous data-scraping pipelines.
— AUG 2025
- Designed and scaled real-time surveillance systems spanning vehicle monitoring, facial recognition, and zero-GPU ANPR.
- Built 3D ergonomic posture-estimation tooling for occupational safety using DeepStream and OpenVINO on constrained edge hardware.
- Developed action-recognition and convex-hull intrusion detection, heavily tuned for CPU-only performance limits.
- Built and optimized detection, segmentation and pose models reaching up to 95% accuracy in real time, and automated data annotation with Vision-Language Models.
— JUL 2023
- Built foundational skills across the data science stack — Hadoop, Spark, and deep learning implementation.
- Delivered end-to-end sentiment analysis and predictive modeling using LSTMs and SQL data foundations.
Technical capabilities.
03 / STACKAgentic AI, RPA & Automation
AI Strategy & Business Analysis
Generative AI & LLMs
Computer Vision
ML Frameworks
Deployment, MLOps & Data
Deployed systems.
04 / SYSTEMSAgentic RPA Platform
IN PROGRESSCustom LLM agents paired with a plugin on the open-source Cua computer-use framework, built to automate repetitive work across departments without the per-bot licensing cost of commercial RPA suites like UiPath.
MCP Tool Layer & LLM Gateway
IN PROGRESSMCP servers exposing ERP and CRM functions — approvals, reporting, reconciliation — as agent tools, served through an LLM gateway for model routing, fallback and cost control, with human-in-the-loop checkpoints and LangSmith evals.
Enterprise Automation Discovery
STRATEGYAI business-analysis engagement for a large enterprise: 600+ processes assessed for automation potential and 100+ process cycles shortlisted for AI and automation rollout, each with scope, effort and ROI.
Hivefloor
MULTI-AGENTLocal-first desktop office for AI coding agents: each agent runs a real CLI in its own terminal, a boss agent routes work by role and availability, with shared memory, Docker sandboxing and approval gates for risky actions.
View on GitHub →SpatialVLM
VLMA grounding head on 4-bit Qwen2-VL-7B that turns "where is it?" into pixel-precise bounding boxes — ~58% Acc@0.5 on RefCOCO, running on an 8GB GPU.
View on GitHub →VLM-Guided Anomaly Detection
VLM · CVUnsupervised defect detection with zero labelled anomalies: CLIP patch embeddings and VLM-written normality priors guide PatchCore coreset selection, benchmarked on MVTec AD and deployable via OpenVINO.
View on GitHub →WebRTC Video Calling
REAL-TIMEP2P browser video calls with Socket.io signalling, plus a GStreamer pipeline that exposes calls as RTSP or YouTube Live for YOLO-based computer vision.
View on GitHub →Transcript → Knowledge Graph
LLM · GRAPHAudio-to-graph pipeline: Whisper transcription, semantic chunking and local-LLM entity extraction (Ollama) into a Neo4j knowledge graph.
View on GitHub →Medical NLP Chatbot
RAGClinical Q&A API with hybrid retrieval — intent routing between SQL facts and ChromaDB vector search — answered by a local LLM.
View on GitHub →SafetyMonitor AI — ANPR
CVZero-GPU ANPR system achieving 180+ FPS on CPU using HOG + PaddleOCR, built for resource-restricted surveillance hardware.
3D Ergonomic Assessment
CVOccupational risk posture analysis using the SMPL body model, scored against REBA/RULA ergonomics frameworks.
Restricted Monitoring AI
CVAction-recognition pipeline using temporal tracking and dynamic convex-hull boundaries to flag facility intrusions in real time.
Traffic Violation Edge AI
CVMobile edge-camera integration running portable YOLO models to log and geo-tag traffic infractions in real time.
Unified Multi-Channel Analytics
CVDockerized platform ingesting multiple real-time RTSP streams for facial ID, crowd tracking, and vehicle classification.
LLaMA Review Intelligence
LLMFine-tuned a LLaMA model on real reviews to generate human-sounding, rating-aware responses, deployed via vLLM.
Fashion Industry RAG
NLPConversational product interface on WhatsApp & Instagram, powered by Claude API and six specialized agents referencing a vectorized retail catalog.
Multilingual Generation System
NLPLocalization pipeline built on Meta's NLLB, paired with autonomous scrapers that transcribe foreign competitive data asynchronously.
Academic background.
05 / EDUCATIONLet's build something
that has to work.
Beyond shipping AI systems, I'm actively looking for a mentor — in this field that's a bit like finding the right co-founder, rare but worth the search. I'm also always glad to connect with people building serious things, or open to guiding someone who is.